Over the past two months, three people I read have buried three different moats. They weren’t writing to each other. Todd Saunders went at it in April, Seema Amble in May, Jaya Gupta this month, each reaching for a different shovel. But line up the eulogies and it’s the same body in every casket.

Saunders went first. Vertical software, he argued, was built on the premise that the vendor understood the customer’s business better than the customer did, and inference erased that premise: a model now absorbs a domain fast enough that nobody’s expertise stays scarce. He sees no middle left. Either you own the rails (payments, identity, compliance, data) and become infrastructure, or you owned only the expertise and become a feature on someone else’s harness.

Amble buried the next one: the interface. When an agent does the clicking, knowing where the buttons live is worth nothing. Strip the screen away and you find out where the value actually lived, which turns out to be the operational logic underneath: the rules and permissions and context an agent needs before it can safely act.

Gupta came at it from the money. Everyone is using AI and everyone is spending, she observed, but usage and spend aren’t a business. The real question is where AI captures the value it creates, and it captures the most where it attaches to budgets, workflows, labor, and accountable outcomes, far more than where it merely shows up.

Three starting points, one direction. What died in every case was the legible: anything you could see clearly enough to measure, a competitor could copy and a model could learn well enough to replace. What survived, in all three eulogies, was the same unglamorous thing none of them had a clean name for: the private, accountable work a model has to be trusted with before it can act.

What Guo Named

This week, gave it one.

Guo, who runs the venture firm Conviction, opened by naming the despair the other three were answering: the 2026 investor’s version of AI psychosis, the conviction that nothing is investable because every company built on a model is a thin wrapper waiting to be absorbed, so buy Nvidia, buy Anthropic, go home. They have the mechanism right, she argues, and the destination wrong.

The mechanism is simple, and she states it more cleanly than anyone: what’s measurable is what’s leaving. Anything you can put on a leaderboard, you can train a model against, and once a model can match it, it commoditizes. It’s why coding agents matured first. A compiler is a free grader and a test suite is a free grader, so you can grind a model against the check until it wins. The receipts are humbling. Mert Demirer and coauthors at MIT studied more than 100,000 developers and found the latest coding agents lifted how much code got written by roughly 180 percent, and how much actually shipped by about 30. Writing got cheap. The judgment about whether a change was right for a decade-old system still runs through a person, because that judgment was never on the test.

So defensibility doesn’t disappear as models improve. It relocates, sliding toward the work a model can’t reach from outside the customer. Guo gives that surviving ground its name, the untrainable corner: frontier work whose correctness exists only inside someone’s private data, walled off inside a system you have to be allowed into. And she locates the real bottleneck precisely. It was never intelligence. You can imagine a model far smarter than any person and it still has to be let in the door, and someone still has to put their name on what it does. The bottleneck is permission, and accountability.

She describes two barriers on that door. The lock is the environment: the security review, the integration, the contract with your name on the outcome. The deadbolt is human habit. A majority of American doctors now open OpenEvidence every day, and no amount of compute buys that. You earn it slowly, on relationships, over years.

Aaron Levie, reading her the same morning, translated the idea into the language of the people deploying it. There’s still an insanely large gulf, he wrote, between what a model can do and what it takes to make it work inside one specific company. That gulf has three parts: technology you build, data you arrange and format, and the customer-by-customer change management of getting an organization to operate differently. The translation never ends.

Which is where Gupta’s money question finds its answer. Value attaches to the places AI has been given permission to sit. Consumer AI throws off enormous surplus and monetizes lightly; enterprise AI touches fewer people and captures far more, because it attaches to budgets, workflows, labor, and measurable outcomes that someone is accountable for.

The Moat That Grows

Here is the turn that should change how you read every AI-eats-the-world headline. Normally a better model is a threat to anyone who isn’t a model company. For the untrainable corner, a better model is a gift.

As capability rises, the line of what’s measurable rises with it, and everything below that line commoditizes and falls away. The value gets pushed up into the shrinking band of work that stays private and accountable. If you’re standing in that band, the next frontier model arrives as a sharper tool you point at the exact workflow you already defend. Your competition gets automated. Your position gets stronger. The moat widens precisely because the models got better.

That reframes the question for anyone building. A model is something your competitor can switch to by Friday. The moat is whose private reality you’re trusted to sit inside, and how much of their accountable, walled-off work you’ve earned the right to touch. I should be transparent that this is the problem space I work on at Typeface, so I come at it with a builder’s bias, and from inside that work I’ve watched Guo’s thesis play out firsthand. I couldn’t agree with her more. The deployments that stick are the ones where we did the slow, unglamorous translation and earned the right to sit inside the workflow, regardless of how clever the model underneath.

It’s also why the convergence landed the way it did. The durable advantage, in every one of these arguments, is the work nobody wants to put on a slide: arranging a company’s messy reality so a model can act on it, earning the security review, holding a skeptical team together through a rebuild that takes quarters. I’ve argued for a while that in enterprise AI the technology is the easier part, and that the people and the process and the plumbing are where deployments live or die. This is that same idea with a balance sheet attached. The unglamorous work is the value.

The sources don’t fully agree on who ends up holding that ground, and the disagreement is the interesting part. Amble thinks incumbents with the deepest data and the heaviest compliance can defend it. Saunders is harsher on anyone who owned only expertise. Gupta is betting on an agent that spans the fragmented systems no single incumbent controls. They split on how you get paid, too. Levie’s answer is to meter intelligently: route the hard work to frontier models, the cheap work to cheap ones, and read a rising token bill as a sign of success. Guo’s answer is to charge for the result. Sierra bills only when its agent resolves a customer’s issue, which works precisely because Sierra owns the private definition of what “resolved” means. One camp optimizes the cost of the tokens. The other makes the tokens someone else’s problem and guarantees the outcome. Those are two very different companies.

What none of them answer is how long any of it holds. Guo is the most honest about the limit. The measurable frontier keeps climbing, which means the untrainable ground is always shrinking under whoever’s standing on it, and you re-underwrite constantly or you slip below the line. So permission might be a durable moat, or it might be an integration lag that standards and data portability eventually erode. Nobody on my reading list this month put a clock on it. That’s the number I want.

The Odd Find

While everyone argued about who owns the enterprise, a team in L. Mahadevan’s lab at Harvard published a quieter result in PNAS: robot swarms move faster when you make them slightly worse at moving. Send the robots along perfectly straight, efficient lines and they pile into each other and seize up. Let them move at pure random and they wander and arrive nowhere. The sweet spot, lead author Lucy Liu found, is a small dose of controlled randomness, a wiggle, that lets them slip past one another and keeps the whole swarm flowing. Too much order turns out to be its own kind of gridlock.

The enterprise has no such trick. You don’t wiggle past a security review, and you can’t randomize your way into a hospital’s trust. The investors’ despair has the mechanism right and the destination wrong. Intelligence keeps getting cheaper, and value keeps sliding toward the few rooms a model has to be let into. Anything you can put on a leaderboard, someone eventually trains against and wins. What lasts is the work that was never legible from outside the room to begin with. The question worth sitting with is not whether your work is hard. It’s whether anyone outside the room can see it well enough to take it from you.

Three conferences opened in the first week of June. Microsoft Build in San Francisco. Snowflake Summit, also in San Francisco. Salesforce Connections in Chicago. Three stages, three keynotes, three sets of slides. And all three companies, without any public coordination, announced the same product category using the same two words: control plane.

Snowflake CEO Sridhar Ramaswamy opened his keynote by positioning Snowflake as “the agentic control plane.” Salesforce had already used the phrase in April when it announced Agent Fabric as a “trusted agent control plane for the multi-vendor AI landscape.” Microsoft Build showcased Agent 365 as a “unified control plane for managing, securing, and governing AI agents.” Box, Ramp, and ServiceNow made structurally identical moves the same week.

Six vendors, one phrase, zero coordination.

Something shifted in the first week of June. And it tells us more about where enterprise AI is heading than any model benchmark.

The Middle Layer

The model layer is commoditizing. That is no longer a controversial claim. GPT-5.5 went generally available in Microsoft Foundry the same week a Chinese open-source model beat it on Humanity’s Last Exam. When the frontier moves that fast, betting on model superiority is betting on a lead that shrinks by the quarter.

What these six vendors are really betting on is the layer above the models: the governance, orchestration, routing, and context infrastructure that determines which agent does what, with whose data, under whose rules. The control plane.

Each vendor arrived at the same architectural conclusion from a different starting point. And the starting point is what makes this interesting.

Salesforce staked its claim on CRM data and MuleSoft’s API fabric. Their Agent Fabric scans and governs agents across Amazon Bedrock, Microsoft Foundry, and any MCP-compatible server. The target is the rules layer: governance and orchestration above the model and the agent runtime. Marc Benioff put it directly at TDX: “Our API is the UI.” The entire Salesforce platform, now exposed as APIs, MCP, and CLI, becomes the substrate that agents operate on.

Microsoft anchored its play in identity and the productivity graph. Agent 365 inherits everything from Entra ID, Microsoft Graph, Teams, and Fabric. If your company already runs on Microsoft, your agents inherit those permissions, those org charts, those data flows. Windows itself is becoming, in Saanya Ojha’s framing, the “DirectX for AI”: a runtime that shapes what agents can see, touch, and modify.

Snowflake built its version on the enterprise data warehouse. Their new Cortex Sense feature automatically builds shared context from the data, business definitions, and operational knowledge that agents need to be useful. CoCo, their autonomous coding agent, now runs on desktop, mobile, Slack, and as a Claude Code plugin. The acquisition of Natoma, an agent governance startup, signals the intent: the data platform itself becomes the trust boundary.

Ramp went vertical with financial data. They just raised $750 million at a $44 billion valuation, tripling in a year, with token spend management as one of the growth drivers. Ramp Stack, launched the day before the raise, is an AI operating system for accounting firms. Geoff Charles, Ramp’s CPO, put it simply: “Firms aren’t asking for another AI tool to prompt. They need something that actually does the work.”

Box chose content and permissions as its moat. Aaron Levie called this “the first year where agents in the enterprise become practical.” Box AI Studio lets admins build custom agents (Legal Reviewer, Brand Steward) per workflow, choosing their own foundation model. Box Automate orchestrates content-centric workflows. When your unstructured data carries permissions, the permission layer becomes the control plane.

And ServiceNow built the kill switch. Their AI Control Tower is an “observe, govern, secure” layer for enterprise agents. Fortune’s headline framed the product pitch as a problem: “Your company’s AI could delete everything in 9 seconds.”

The Data Plane

Here is the pattern worth naming. Every vendor’s control plane maps to the data asset they already own.

Salesforce owns the customer record. Microsoft owns the identity graph. Snowflake owns the analytical data. Ramp owns the spend ledger. Box owns the content corpus. ServiceNow owns the IT workflow. Each of them is making the same argument: the data gravity well you already orbit is where the agents should live.

This is a classic platform dynamics play, and the June conferences weren’t even the first move. Google Cloud made a structurally identical bet at Next in April, launching the Gemini Enterprise Agent Platform with Agent Identity, Agent Gateway, and Agent Registry as core governance infrastructure. Bain published an analysis of that event titled “The Agentic Enterprise Control Plane Comes into View.” Their thesis: the company that owns the control plane owns the customer relationship for the next decade. Futurum called it “one of the industry’s most aggressive attempts to establish a universal enterprise AI control plane.”

The numbers beneath the pitch are sobering. Salesforce’s 2026 Connectivity Report found organizations now average 12 agents, and half of them operate in silos. Gartner projects 40% of agentic AI projects will fail by 2027 because governance hasn’t kept pace.

I build content infrastructure at Typeface, where the interaction between AI agents and enterprise data is the whole product surface. What I see from the inside tracks what these announcements reveal from the outside: the constraint on enterprise AI is organizational trust infrastructure. Who can see what, who approved what, and who’s accountable when the agent does something wrong.

I wrote earlier this year about permission lock-in: the idea that AI vendor lock-in lives in the accumulated decisions, permissions, and organizational knowledge embedded in how you use the platform. Every control plane announcement this week validates that thesis. The lock-in is the governance layer you can’t migrate.

The Talent Plane

If you’re a product or marketing leader reading this, the control plane war is also a talent war. Every vendor can ship the governance layer. The people who can wire it into an actual organization are scarce.

The control plane thesis has a gap. It assumes the organizational plumbing is ready. Most of it isn’t. As one practitioner put it this week, the operating model is the bottleneck, and missing from the agentic literature are the power dynamics and trust workstreams that actually live on leaders’ calendars. You can build the control plane. If the org chart doesn’t match the agent topology, the control plane governs nothing.

Andrew Ng sees the same gap from the hiring side. The Forward Deployed Engineer role, pioneered by Palantir and now resurgent at OpenAI and Anthropic, exists because someone needs to bridge the gap between the platform and the organization. Ng predicts the role will fragment into LLMOps, Evals, and Harness Engineers. The people who understand both the technology layer and the organizational layer, who can translate agent topology into org design, are about to become the most valuable hires in enterprise software.

Salesforce’s own engineering philosophy reflects this tension. They call it “guided determinism”: fixed handoff rules between agents with LLM reasoning operating within those guardrails. The deterministic layer sits on the outside. The probabilistic intelligence operates within. It’s an explicit admission that fully autonomous multi-agent orchestration isn’t enterprise-ready. The humans set the rails. The agents ride them.

The multiplayer problem I wrote about last month is the same problem in different clothes. Individual AI productivity is real. Organizational compounding requires shared infrastructure. And shared infrastructure requires something no vendor can ship: agreement about who’s in charge, what matters, and which agents get to act without asking. That’s an org design problem wearing a technology costume.

The Odd Find

A Numerator study on consumer AI adoption found something that breaks every model we have for technology diffusion. A 25-year-old and a 65-year-old are equally likely to sit anywhere on the AI adoption spectrum. PCs, social media, and smartphones all skewed young. AI is the first major technology in decades where age doesn’t predict adoption tier. The early-adopter playbook that governed technology marketing since the 1990s may not apply to the thing everyone assumed would follow it most closely.

Six vendors in one week, each claiming the control plane, each building it on top of the data they already own. The convergence is real. The question it raises is whether any single vendor can govern a landscape where the agents span all six data gravity wells at once.

The answer, most likely, is that they can’t. And the organizations in the middle, the ones running 12 agents across 4 vendors, will spend the next two years building the governance layer themselves. Not because the platforms failed. Because the platforms all succeeded, in six different directions at once.

Coach grew 29% last year on a bet that would have gotten most CMOs fired: human relevance over performance optimization, in a year when every other enterprise was chasing the opposite.

Joon Silverstein, Coach’s global head of marketing, laid out the numbers at Brandcast NYC: 15 million organic engagements, 450,000 user-generated posts, a 60% awareness lift among Gen Z. Stop treating brand as the top of the funnel. Brand is the funnel. Relevance-first, while every competitor doubled down on attribution models and AI-powered optimization.

Meanwhile, a Gartner study of 350 executives at billion-dollar companies found that 80% of organizations deploying autonomous AI have already cut their workforce. The layoffs aren’t delivering returns: reduction rates are nearly identical among companies reporting high ROI and those seeing none.

One side of the market is rewarding human conviction. The other side is punishing human hesitation, in the same year, in the same economy. The contradiction is a sorting.

The Human Quotient

Human understanding, judgment, creativity, and trust are becoming even more valuable, not less. That’s the central thesis of the Marketing2030 report, released this week by the Institute for Real Growth and Oxford’s Said Business School.

They call it the Human Quotient: the compound of human and AI, where human capabilities serve as the multiplier. The report, built from interviews with global business leaders, argues marketing is being redefined from a communications function to an enterprise capability centered on understanding people, culture, trust, and value creation.

If you buy that framing, the CMO role doesn’t shrink. It becomes the most consequential function in the building.

Eugene Healey would push back on the timing. He documents the structural instability: the CMO already has the shortest C-suite tenure, a 54% misalignment rate with their board, and a growing number of systems-thinking leaders from other functions circling the job. His “barbell model” predicts surviving CMOs will need both infrastructure design and brand instinct, while the execution middle gets commoditized. The role endures. The version that survives might be unrecognizable.

Coach’s numbers suggest Healey’s barbell is already forming. Silverstein didn’t optimize the execution layer. She rebuilt the brand thesis around relevance and let the execution follow. The 450,000 UGC posts weren’t manufactured. They happened because the brand said something worth repeating. Performance marketing would have bought impressions. Taste earned participation.

I build content infrastructure at Typeface, so the line between “human” and “automated” work is something I think about constantly. What Coach demonstrates is that the line sits higher than most organizations draw it. Production is table stakes: write the brief, make the asset, hit the calendar. The human premium is taste and judgment. Knowing what the brand stands for, and having the nerve to bet on it when every dashboard rewards the opposite.

The Walkaway

The market is telling us where the premium sits. At every layer, people are walking away from organizations that won’t listen.

Fortune reported that 38% of job candidates walked away from a hiring process because it required an AI-conducted interview. The technology worked fine. The candidate decided the company wasn’t worth talking to if the company wouldn’t talk to them. 70% were never told upfront that AI would be evaluating them. The tool designed to scale the hiring funnel is actively shrinking it.

ManpowerGroup’s Global Talent Barometer tells the same story from inside organizations. AI usage climbed to 45% while confidence dropped 18%. A behavior they call “job hugging” is taking hold: 64% of workers are staying put at their current employer, seeking stability rather than opportunity. Loyalty has nothing to do with it. They’re hedging. The signature of a system that adopted the tools without earning the trust.

Ethan Mollick, in “Choosing to Stay Human”, names the deeper choice. AI handles delegation well enough now that the tempting path is to hand everything off. The harder path is to stay in the loop: to use AI for learning and augmentation rather than replacement. The people who delegate everything move faster in the short term. The people who stay engaged keep developing the judgment that makes them irreplaceable.

The irony is structural. The skill AI can’t replicate (human judgment) atrophies fastest when you let AI replicate everything else. Andrej Karpathy has been saying it more bluntly: you can outsource your thinking, but you can’t outsource your understanding. The people who hand everything off outsource both.

The walkaway takes different forms: candidates leaving AI interviews, workers hugging their tasks, consumers whose enthusiasm for AI-generated content collapsed from 60% to 26% according to a Billion Dollar Boy study covered by eMarketer. Call it what it is: a market signal. People are telling organizations, in every language available to them, that the human parts matter. And organizations keep measuring something else.

If you’re a marketing leader reading these numbers, the real question is which decisions you protect from AI.

I wrote this week about the three problems nobody told you about enterprise AI: ghost systems, rubber-stamping, and beautiful empty rooms. Those problems are what happens when the human premium gets dismissed as friction. Ghost systems run without human judgment. Rubber-stamping pretends humans are in the loop when they’re not. Empty rooms mean the infrastructure works but nobody cares about what it produces.

The Human Quotient is what fills those rooms. People who still care whether the output matters, and organizations willing to treat that caring as an asset rather than a cost line.

The Odd Find

Duolingo killed their mascot last year. Duo, the green owl that had terrorized millions into finishing their Spanish lessons, was announced dead across every social channel, hit by a Cybertruck. The results are now fully tallied: 1.7 billion impressions in two weeks, monthly Android users up 25%, downloads up 38%. To bring Duo back, users had to collectively earn 50 billion XP across 15 countries. They hit the target. The most effective growth campaign of the year involved no attribution models, no AI optimization, and no performance marketing. Just a dead cartoon owl and the human impulse to bring it back.

In a magic trick, there are three acts. The pledge is the setup: here is something ordinary. The turn makes it disappear. The prestige brings it back.

Every organization in 2026 has completed the turn. The human vanished from the workflow, the hiring process, the review queue, the content calendar. Coach grew 29%. Candidates walked away at 38%. Fifty billion XP to resurrect a cartoon owl.

The question for the rest of this year isn’t whether AI works. It’s whether anyone knows how to do the prestige.

There is a confession I hear in almost every conversation with a marketing leader. It surfaces sideways, usually after the second or third meeting, once the vendor pitch is over and we’re talking honestly. “We built the machine. And now we can’t stop feeding it.”

The details vary. Sometimes it’s the monthly content calendar with 47 line items that nobody reads. Sometimes it’s the review process that takes longer than the writing. Or the agency relationship where the briefs became the product and the content became filler. One leader told me her team went from writing ten blog posts a quarter to sixty, and not a single one outperformed the original ten. The specifics change. The exhaustion is always the same.

I build content infrastructure at Typeface, so I hear these confessions more than most. But what struck me this week wasn’t any single conversation. It was that two people I’d never connected before, working in completely different contexts, described the same structural trap from the inside. And the way they’re escaping it tells us something about where marketing is right now, beneath the AI adoption numbers.

Like-Minded Lunatics

Julianne DeVincenzo runs global content strategy at Optimizely. In an interview with State of Brand, she described what she inherited with a line I can’t stop thinking about: “We spent so many years feeding the machine that we forgot we ever loved the work.”

Her diagnosis is precise. Content teams became order-takers operating what she calls the “butcher counter model,” where everyone takes a number and waits for their request to be filled. The strategy served the machine rather than the other way around. High volume, low coherence, no one asking whether the output mattered.

DeVincenzo didn’t reorganize from the top down. She found what she calls “like-minded lunatics,” people already in pain from existing systems, and built a coalition from the middle. It took seven months. Sales teams joined unexpectedly, recognizing the same dysfunction from their side. The mandate that emerged: make the team “structurally smaller and dramatically more capable.” Subject matter experts now communicate directly rather than routing everything through marketing approval. The machine finally serves the strategy.

At AirOps, Jess Rosenberg faced a different version of the same problem when building Quill, their AI marketing agent. The default AI voice was available out of the box: that eager, hedge-stacking, congratulatory tone that sounds like every chatbot you’ve ever closed. Her team rejected it. They built an archetype they called the “Expert Colleague”: a mentor with earned authority. Warm, confident, direct, occasionally surprising.

The banned-patterns list proved as important as the voice principles themselves. They wrote it as carefully as the positive guidelines: no hollow affirmations, no hedge-stacking, no congratulatory openers, no “aren’t just X, they’re Y” constructions. The entire voice system was delivered to engineering as a markdown file. Rosenberg’s thesis: “Capability without character is forgettable.” Domain expertise earns the right to have personality. Everything else is wallpaper.

DeVincenzo told one more story. Her AI tools kept flagging em dashes as unnecessary. She overruled them. Not because em dashes are sacred, but because letting an algorithm make your stylistic choices is where voice starts to die. (I ban em dashes in my own writing, so the irony is not lost on me.) Her reasoning: “Taste cannot be automated.”

What the Dashboards Miss

Taste cannot be automated. And the data is starting to prove it.

DeVincenzo is not alone in the diagnosis. A BuildBetter survey surfaced a number that stopped me cold: 84% of product leaders claim AI is integrated across their lifecycle. Only 18% of the managers who do the work agree. That’s a 4.7x perception gap between the people who approve the dashboards and the people who live inside the workflows.

The machine looks different from the top than it does from the middle. And that gap matters because the people who decide whether to deploy AI across content operations are often the same people most insulated from the experience of using it. The dashboard says integrated. The workflow says otherwise. DeVincenzo saw this firsthand: the executives who approved the content calendar had no idea what it felt like to fill it.

On the consumer side, the antibodies are forming fast. According to a recent study tracking consumer attitudes toward AI-generated content, enthusiasm collapsed from 60% to 26% in roughly two years, with young adults (22-34) detecting AI writing at nearly 89%.

Volume is cheap. Trust is expensive. And trust is the thing you can’t scale by adding more tokens.

If you’re a marketing leader reading those numbers, the implication is uncomfortable. The same AI tools that gave your team a productivity bump are simultaneously training your audience to distrust the output. The speed gain is real. The credibility cost is real too. The question is which one compounds faster.

DeVincenzo didn’t reject AI. She used it for research, ideation, and structural work, then kept humans in control of editorial judgment. The restraint was the strategy. You could call it taste. You could also call it knowing which decisions to protect from the machine you built.

I wrote a few weeks ago about the multiplayer problem in AI productivity: the gap between individual speed and organizational compounding. What DeVincenzo and Rosenberg are showing is the layer beneath that problem. Before you can build shared infrastructure, someone has to remember what the infrastructure is for. Coordination assumes shared purpose. These practitioners had to rebuild the purpose first.

Everyone I talk to agrees the machine is broken. The interesting divergence is in what they do next. Most organizations respond by adding another layer: a governance framework, a brand guide, a review stage, a new tool on top of the old tools. DeVincenzo and Rosenberg went in a different direction. They asked who still cares, and they gave those people authority.

I’ve been writing about content operating systems for months. The thesis has always been structural: the real leverage sits in the infrastructure layer, in shared platforms, in the organizational compounding that individual tools can’t deliver alone. I still believe that framing is right. But this week’s convergence sharpened something I’d been underweighting.

Infrastructure without conviction is just plumbing. You can build the platform, wire the knowledge graph, close the measurement loop. And if the people inside the system have forgotten why the work mattered in the first place, the infrastructure automates the emptiness. The system runs faster, but it’s running nowhere.

The Odd Find

Spotify turned 20 this month and celebrated by replacing its logo with a disco ball. For five days. The backlash was immediate and loud, which was exactly the point. Every complaint was free distribution. A five-day deviation from a decade-old brand identity generated more conversation than most companies achieve in a year. They didn’t add a feature or launch a product. They temporarily broke the most recognizable thing they had, on purpose, and let the breakage do the work.

DeVincenzo’s “like-minded lunatics” aren’t a change management tactic. They’re a taste test. The question she’s really asking: does anyone here still care enough to fight the machine we built?

The machine is always hungry. The question is whether you remember what you were trying to cook.

Foraging: Life Moves Pretty Fast

ActivTrak tracked 443 million hours of work. AI made everything faster. Including all the wrong things.

Something shifted in the conversations I’ve been having this month. Every leader says the same thing: we’re moving faster than ever. The dashboards confirm it. And almost nobody can explain what, exactly, got better.

I think speed became the default metric for the AI era. And I think it’s exactly the wrong one.

The Acceleration Paradox

There is a specific kind of fatigue that comes from running harder without gaining ground. It has a shape, and the data is starting to trace it.

The pattern I see in almost every marketing org I talk to: the content team is 3x faster. The approval chain takes just as long. The campaign cycle hasn’t budged. Individual speed went through the roof. Organizational speed barely moved.

Alexander Atzberger calls this the iceberg problem. The tip is visible: content that took days now takes minutes. Below the waterline, handoffs and cycle times are unchanged. Gartner predicts 40%+ of agentic AI projects will be canceled by end of 2027, largely because organizations solve the tip and call it transformation.

Microsoft’s Work Trend Index surveyed 20,000 workers and found something that should worry anyone betting on tools alone: organizational factors drive 2x greater AI impact than individual skill. The report maps a progression from using AI for individual tasks to restructuring entire workflows around it. Most companies haven’t made it past the first step. The tools outpaced the org charts.

The fracture is now visible at the top. A Workplace Intelligence study of 2,400 knowledge workers found 54% of C-suite say AI is “tearing their company apart.” 75% admit their strategy is “more for show.” Nearly all of them, 97%, have deployed AI agents in the past year. And 60% are planning layoffs for employees who don’t adopt.

The message is clarifying: move faster, or move out. But when 79% of those same organizations report challenges in making AI work, the speed demand starts to look like a different kind of dysfunction. These are companies running at two speeds, with leadership measuring one and experiencing the other.

Aditya Agarwal, who scaled engineering at Facebook and served as CTO of Dropbox, named the deeper risk: good speed compresses learning cycles, bad speed optimizes for novelty over depth. AI makes novelty trivially easy. The edge is choosing what is worth building and staying long enough to learn something the market doesn’t know yet. Lower friction can also mean accelerating false starts.

If you’re a marketing leader reading this, ask yourself: is your team faster at the things that compound, or faster at the things that expire?

Where the Work Piled Up

Speed up one part of any system and you expose the next constraint. Every engineer knows this intuitively. Most organizations are learning it the hard way.

The clearest example is code review. AI-assisted engineering teams merge 98% more pull requests, according to a Faros AI analysis. And code review time increased 91%. Developers report higher productivity and more time spent reviewing other people’s code. Judgment became the bottleneck.

This is Amdahl’s Law at organizational scale. The principle is simple: the overall speed of a system is limited by the part you can’t parallelize. You parallelized the writing. You exposed the understanding. The review layer, the part that decides whether the code is correct and architecturally sound, doesn’t compress the same way. The humans doing the reviewing are the fixed-speed stage.

A survey of 900+ engineers by the Pragmatic Engineer confirms the practitioner experience: “shippers,” the engineers who benefit most from AI, also accumulate tech debt faster. They write more, commit more, and leave more for the next person to untangle. Engineering managers and individual contributors are converging on the same work, because when AI writes the code, the remaining work is judgment and review. Companies spend $100-200 per month per engineer on AI tools. Nobody has figured out how to fund the review capacity those tools create.

And the spending is running ahead of the strategy. Ramp’s May 2026 AI Index shows Anthropic passing OpenAI in business adoption for the first time: 34.4% vs. 32.3%. Anthropic quadrupled year-over-year while OpenAI grew 0.3%. Companies are buying faster than they can absorb. Separately, Uber reportedly burned through its entire 2026 AI budget by mid-year. If a company with that engineering bench and that operational discipline can’t manage AI spend, the rest of the market should pay attention.

When Fast Stops Being a Moat

The bottleneck migration in engineering is a preview of something larger. When AI compresses the building, what happens to the people whose advantage was building faster?

Two enterprise deals died the same way, as Gokul Rajaram documented in late April. Buyers chose to build internally rather than renew. Above $500K in annual contract value, a domain expert with Claude Code can replicate a functional workflow tool in weeks. It won’t have the polish. It won’t scale the same. But “good enough” is all procurement needs when the alternative is writing another six-figure check. The vendor’s speed advantage, the thing that justified the contract, evaporated.

This is the acceleration paradox applied to markets. SaaS companies built moats on being faster than their customers. AI collapsed that gap. The gap between “buy” and “build” that was months or years is now weeks. Enterprise teams carry AI transformation mandates on 2026 OKRs. Replacing a vendor with an internal build is a two-for-one: kill the budget line and claim the AI transformation win. The switching cost isn’t worth the hassle at $50K. At $500K+, the VP has a self-writing business case. The vendor’s moat didn’t leak. It was never as deep as the contract implied.

So if speed isn’t a moat for the individual worker, isn’t a moat for the engineering team, and isn’t a moat for the software company, what is?

Dan Rosenthal offers an answer. He scaled from 2 to 20 people in 9 months by dogfooding a closed-loop go-to-market system: Clay, HubSpot, and AI agents forming a seven-step pipeline where step 7’s performance data refines step 1’s targeting. Every outbound sequence teaches the next one. Every reply rate adjusts the next round of personalization. The system compounds because each cycle sharpens what comes after. It didn’t start fast. It got fast by staying in the loop long enough to learn.

That is a different kind of fast. Not faster at executing. Faster at learning. Rosenthal’s system compounds because every output feeds back into the inputs. Atzberger calls this “multi-player AI”: what one person discovers becomes available to everyone. Microsoft’s research describes the same pattern at enterprise scale, where human oversight narrows to the decisions that matter and agents handle the rest. The design is what creates the speed.

The organizations that treat AI as an accelerant for the existing model end up producing more of everything and improving nothing: more email, more chat, more weekend hours, same outcomes. ActivTrak’s data is the proof. The ones that redesign the model, that ask which work should exist at all, end up with systems that compound. And compounding is a property of design.

The Odd Find

Meanwhile, in a corner of the internet unrelated to AI productivity: Warburtons, a 150-year-old British bakery, convinced Transport for London to rename Baker Street station to “Bakers Street” for two days. They replaced platform announcements with “mind the bap” and “stand behind the buttery yellow line.” A bread company hijacked the London Underground. No algorithm, no agent, no dashboard. Just a good joke and whatever internal approval chain greenlit “mind the bap.” Sometimes the fastest path to attention is not speed at all.

If you’re in a room this week where everyone is celebrating velocity, it might be worth asking the follow-up: velocity toward what?

TL;DR

• Meta, Microsoft, and Salesforce all incentivized AI token consumption. The result: “tokenmaxxing,” developers gaming volume metrics the same way they once gamed lines of code. Shopify quietly got it right by pairing usage data with output quality.

• The Federal Reserve compared three surveys measuring U.S. AI adoption and got three answers: 18%, 41%, and 78%. The divergence is structural. Every AI stat you’ve cited carries embedded assumptions about who counts and what counts.

• Microsoft surveyed 20,000 workers and found organizational factors drive 2x more AI impact than individual skill. But only 13% of employees are rewarded for reinventing how they work. The incentive structure contradicts the transformation goal.

• PwC found 75% of AI’s economic returns captured by 20% of companies, with leaders outperforming by 7.2x. The gap is widening. Catching up requires measuring operational outcomes.

The reports that stacked up this week told the same story from different angles. The metrics organizations use to track AI adoption have almost nothing in common with the metrics that predict AI value.

Goodhart’s Token

Meta built an internal leaderboard ranking all 85,000+ employees by token usage across its AI coding tools. Tiers included “Session Immortal” and “Token Legend.” The company consumed 60.2 trillion tokens in 30 days, which would run about $900 million at standard API rates. Meta discontinued the leaderboard after media exposure. An engineer suggested the real purpose was generating training data for Meta’s coding models.

They were not alone. Microsoft incentivized token consumption through internal dashboards starting in January. A Windows division engineer admitted gaming the metrics: reprocessing documentation that didn’t need reprocessing, prototyping features never meant to ship, running agents inefficiently. The fear of being labeled insufficiently “AI-native” was enough. Salesforce set minimum monthly spend targets ($100 for Claude Code, $70 for Cursor) and recently removed spending caps entirely. Developers request throwaway projects solely to burn tokens.

This is Goodhart’s Law playing out in real time. “When a measure becomes a target, it ceases to be a good measure.” Token consumption is the new lines of code. It tells you something is happening, but nothing about whether that something matters.

Shopify found a quieter path. They celebrated high usage only when paired with great work. They rebranded their internal leaderboard as a “usage dashboard” to discourage competition. They added circuit breakers for runaway agents. The question they optimized for was not how many tokens were consumed, but which tokens cost the most relative to what they produced.

Three Surveys, One Economy, Three Answers

The Federal Reserve published a paper in April comparing three federal surveys that all measure AI adoption in the U.S. economy. The results:

• BTOS (Business Trends): 18% of firms

• RPS (Real-Time Population): 41% of workers

• SBU (Business Uncertainty): 78% of workers at AI-adopting firms

All three describe the same economy at roughly the same time.

The spread is structural. Each survey asks a different question to a different population with different sampling weights. BTOS asks about “producing goods and services,” a narrow definition. SBU asks about “business functions,” which is broad. Question framing alone accounts for a significant portion of the gap. Social desirability bias adds another layer: executives face pressure to demonstrate AI initiatives, inflating top-down estimates relative to bottom-up worker reports.

And even within that 41% who use AI, the engagement is shallow. Only 12% use it daily. The modal GenAI worker touches it a few times a week for less than an hour. The distance between “41% use GenAI at work” and “12% use it daily” is the distance between adoption and fluency.

Behind the headline numbers, sector variation tells a sharper story. Financial services: 63% GenAI usage. Accommodation and food services: roughly 3%. A 20:1 ratio between the highest and lowest adoption sectors. AI is concentrating in knowledge-work sectors, and the concentration is accelerating.

Microsoft’s Global AI Diffusion report adds the international layer. 17.8% of the global working-age population now uses generative AI. The distribution is strikingly uneven: UAE at 70.1%, the US ranking 21st at 31.3%, the Global South averaging 15.4%. The most counterintuitive data point: git pushes increased 78% year-over-year globally, yet US developer employment grew 8.5% to 2.2 million. More AI, more developers. The Jevons paradox at work.

The Org Is the Bottleneck

Microsoft surveyed 20,000 workers across 10 countries and found that organizational factors (culture, manager support, talent practices) account for 2x more AI impact than individual factors. The ratio: 67% organizational vs. 32% individual. Training employees on AI tools is necessary. It’s also insufficient. The constraint is how work is structured around people.

The study defines a maturity spectrum from Author (you produce work, AI assists on specific elements) through Editor and Director to Orchestrator (multiple agents running parallel workflows with human oversight on exceptions). Most organizations are stuck between Author and Editor. Getting to Director and Orchestrator requires restructuring jobs, teams, and incentive systems.

The incentive gap is where this gets uncomfortable. 58% of AI users say they produce work they couldn’t have created a year ago. But only 13% receive rewards for reinventing how they work with AI. Meanwhile, 45% feel safer sticking to current goals. The system rewards performing over transforming.

Brian Armstrong decided to cut through the ambiguity. Coinbase laid off 14% of its workforce, roughly 700 people, and announced a restructuring he framed as “rebuilding Coinbase as an intelligence, with humans around the edge aligning it.” Pure management roles eliminated. Every leader an active individual contributor. Hierarchy capped at five layers. The most radical move: one-person AI pods where a single individual handles engineering, design, and product management backed by agent fleets. Where Microsoft published a maturity spectrum, Armstrong is executing the endpoint.

Measuring What Compounds

Every.to runs five software products with 15 people and 100% AI-written code. Each product is primarily built and run by one person. Their measurement reframe: target real work outcomes like “reduce onboarding task time by 40%.” When execution gets cheap, strategy and taste become the scarce resources.

PwC’s global study from early April surveyed 1,217 senior executives across 25 sectors and found 75% of AI’s economic gains captured by 20% of companies. Leaders outperform peers by 7.2x. The differentiator: leaders treat AI as a self-optimizing operational layer that continuously adapts. They’re making decisions without human intervention at 2.8x the rate of peers. The gap is widening.

Capital keeps flowing toward this bet. Deedy Das’s running list of “Neolabs,” pre-revenue AI startups at $1B+ valuations, grew from 50 to 63 in a single quarter. Redpoint’s CIO survey from late March found 46% of enterprise CIOs open to AI-native startups over incumbents. Swyx’s read: “CIOs are more hungry than conservative right now, and that will not last.” The window is real but temporal.

The odd find

Snapchat is putting brand AI agents directly into users’ chat tabs.

AI Sponsored Snaps let users have actual conversations with brand-operated agents: ask a question about a product, get a personalized recommendation, tap to buy. Early data shows 22% more conversions than standard ads. The surface area is staggering. Snapchat users sent 950 billion messages in Q1 2026, roughly 10.5 billion a day. Brands can now inject conversational agents into that stream.

The ad unit is no longer an impression. It’s a conversation. And that breaks every measurement framework the industry has built over the past two decades. CPM, CTR, ROAS: all assume a human saw something static. When the “ad” is an AI agent having a back-and-forth dialogue, what exactly are you measuring? Conversations started? Questions answered? Minutes of brand engagement? Nobody has agreed yet, which means the companies deploying these agents are flying on early conversion data while the measurement infrastructure catches up.

This is where the measurement theme of this entire edition lands in marketing specifically. The old metrics don’t describe what’s happening. The new ones don’t exist yet.

What I’m Thinking About

Earlier this week at the Lightspeed fireside, Raviraj Jain said something that stayed with me: “We are all managers of agents.” I’ve been thinking about what that means for measurement.

If you measure how many agents your team runs, you get Meta’s leaderboard. If you measure what those agents produce, you get Shopify’s dashboard. If you measure how the organization restructures around agents, you get Microsoft’s Author-to-Orchestrator spectrum.

The pattern I keep seeing in customer conversations is similar. (I should be transparent: this is the problem space I work on at Typeface.) The moment you solve content creation with AI, the bottleneck migrates to content approval. Solve approval, it moves to distribution. Solve distribution, it moves to personalization. The companies clearing PwC’s 75/20 threshold are the ones measuring their distance to the next bottleneck.

The metric that compounds is bottlenecks eliminated.

Sources:

• Gergely Orosz, “Tokenmaxxing”: AI Usage Metrics Gone Wrong | The Pragmatic Engineer, April 2026

• Federal Reserve: Monitoring AI Adoption in the U.S. Economy | FEDS Notes, April 2026

• Microsoft: The State of Global AI Diffusion in 2026 | Microsoft, May 2026

• Microsoft: How Frontier Firms Are Rebuilding the Operating Model | Microsoft, May 2026

• Brian Armstrong: Coinbase AI-Native Restructuring | X, May 2026

• Every.to: AI Product Management Guide | Every, May 2026

• PwC: 2026 AI Performance Study | PwC, April 2026

• Deedy Das: The Ultimate List of AI Neolabs | X, May 2026

• Shawn Wang (swyx): Redpoint SaaS Redo List | X, March 2026

• Lightspeed Venture Partners: The Next Era of Marketing panel | May 2026

• Snapchat: AI Sponsored Snaps | Snap Newsroom, May 2026

My prior writing on these topics:

• The Agent Experience Gap | Substack, April 2026

• The Three Adoption Playbooks | Substack, April 2026

TL;DR

• Karpathy declared vibe coding over and replaced it with “agentic engineering.” But the bigger reveal was Software 3.0: his entire MenuGen app became unnecessary when a single prompt to Gemini did the same thing directly in the pixels. Programming is now prompting.

• Todd Saunders extended this to vertical SaaS: when AI can learn any domain, companies split into “rails companies” (infrastructure) or features on someone else’s harness. “There is no third outcome.”

• While the paradigm shifts, the productivity numbers everyone cited were wrong. Karpathy cited METR and Stanford: the real median gain is 10-15%, not 60%. It takes 30-100 hours of deliberate practice before the tools consistently pay off.

• ActivTrak found only 3% of workers are in the productivity sweet spot (7-10% of time in AI tools). 83% of workers in a UC Berkeley study said AI increased their workload. The gap between adoption and effective use is 19x.

• Expedia bet its Gen Z strategy on a livestreamer. 25% of Gen Z bypasses travel sites entirely. Expedia built a content-to-commerce pipeline with IShowSpeed (150M followers): watch the trip, book the trip. The booking app isn’t the product. The content surface is. Same logic as MenuGen, different industry.

Karpathy vibe-coded a restaurant menu app. Full stack: OCR pipeline, Vercel deployment, image generation. A real product. Then he saw the Software 3.0 version: take a photo of the menu, give it to Gemini, say “overlay images onto the food items.” The model rendered results directly into the pixels.

“All of my MenuGen is spurious. That app shouldn’t exist.”

That sentence landed differently than most conference quotes. A builder looked at something he built and realized the entire category of effort was unnecessary. Not because the app was bad. Because the paradigm shifted underneath it.

The paradigm

At Sequoia’s AI Ascent, one year after coining “vibe coding,” Karpathy killed his own term. His replacement, “agentic engineering,” gets the headlines. But the deeper shift is the one he almost buried in the middle of the talk.

Software 1.0: writing code. Software 2.0: arranging datasets, training neural nets. Software 3.0: prompting. The context window is the lever. The LLM is the interpreter. Programming is now prompting.

This isn’t code getting faster to write. It’s a new category of information processing that wasn’t possible before. LLM knowledge bases are another example Karpathy gave: “This is not even a program. There was no code that would create a knowledge base based on a bunch of facts.”

Vibe coding raised the floor (everyone can build software). Agentic engineering preserves the ceiling (quality, security, reliability). They are different things, not sequential stages. The developer keeps taste, judgment, and specification. Agents handle implementation.

What this eats

If Software 3.0 means the context window replaces the codebase, entire categories of applications become unnecessary. Karpathy’s MenuGen is the small version. The large version is playing out across industries.

Todd Saunders extended the logic to vertical SaaS. When AI can be trained on any domain quickly, proprietary domain knowledge stops being a durable moat. Companies split into “rails companies” (owning payments, identity, compliance, data as infrastructure) or “domain-only companies” that become features on someone else’s harness. “There is no third outcome.”

Ben Lang published YC’s emerging playbook for AI-native companies: “token-max, not headcount-max” as the operating metric. Software factories where repos have no handwritten code, just specs and test harnesses where agents iterate until tests pass. Every action must produce an artifact AI can learn from to make the company “queryable.”

In February, Azeem Azhar and Nathan Warren provided the macro data anchoring this: monthly AI revenue grew from $772M (January 2024) to $13.8B (December 2025). 18x in 24 months. Realized revenue, not projected TAM.

The capital is following the paradigm. The question is whether organizations can.

Meanwhile, on the ground

They mostly can’t. Not yet.

Gallup crossed 50% this week. Half of all U.S. employees now use AI at work. Morgan Stanley, in the same week, found that regular AI usage jumped 13% while confidence in using the technology fell 18%. The people who use AI most are the least confident they understand it.

Karpathy quantified why. At the same Sequoia talk, he cited the research everyone should have been reading: the METR study found experienced developers on their own codebases worked 19% slower with AI tools. Stanford measured median productivity gains at 10-15%, not the 60% that went viral. And it takes 30-100 hours of deliberate practice before the tools consistently pay off.

ActivTrak analyzed real-time productivity data and found the sweet spot at 7-10% of time in AI tools. Only 3% of workers are there. 57% spend less than 1%. An eight-month UC Berkeley ethnographic study found that 83% of workers said AI increased their workload: expanded task scope, dissolved boundaries between work and rest, parallel processing until the aggregate effect was exhaustion.

Korn Ferry found the leadership gap underneath: 52% of organizations plan autonomous AI agents, but only 22% believe their leaders can manage human-AI teams. The paradigm is shifting. The organizations aren’t ready.

Four voices described different facets of the same gap. In an internal memo that leaked last April, Shopify CEO Tobi Lutke made AI usage a structural condition of employment: “Reflexive AI usage is now a baseline expectation.” Gokul Rajaram predicted product design will cease to exist as an independent function by end of 2026. Jaya Gupta named the structural mechanism: experience is now a tax. And Aaron Levie predicted a consulting onslaught because “there is no shortcut to change management for the enterprise.”

The odd find

Software 3.0 isn’t just eating apps. It’s eating how people discover things.

25% of Gen Z bypasses online travel agencies entirely. They decide where to travel on TikTok and YouTube, not on brand domains. Expedia’s response: a year-long partnership with IShowSpeed (150 million followers), structured as a full content-to-commerce pipeline.

They built a TikTok account (@Exspeedia_) for real-time clips from Speed’s travels, and a custom microsite (Exspeedia.com) where viewers book the exact flights, hotels, and activities featured in the content. The gap between inspiration and transaction collapses to zero.

This is Karpathy’s MenuGen logic in a different domain. The booking app isn’t the product. The content surface is. The traditional interface, a search box on a travel site, becomes as unnecessary as Karpathy’s OCR pipeline once you realize the user already decided what they wanted before they ever opened the app.

What I’m thinking about after this week’s reading

Karpathy looked at his own app and saw it was unnecessary. That’s the cleanest version of where we are. The paradigm is shifting from building software to prompting intelligence, and the shift is moving faster than most organizations can absorb it.

The 30-100 hour practice threshold is the most useful number in all of this. It connects the sweet spot (3%), the confidence paradox (usage up, confidence down), and the leadership gap in a single frame: the technology moved from 1.0 to 3.0 while the org chart is still debating 2.0.

The organizations that will cross the gap are the ones that budget for the learning curve instead of expecting the tools to work on contact. The paradigm rewards practice, not enthusiasm.

Sources:

• Andrej Karpathy at Sequoia AI Ascent: “From Vibe Coding to Agentic Engineering” — YouTube, April 2026

• ActivTrak: AI Adoption Trends and Workforce Productivity — ActivTrak Research, April 2026

• HBR: AI Doesn’t Reduce Work — It Intensifies It — Harvard Business Review, February 2026

• Korn Ferry: TA Trends 2026 — Human-AI Power Couple — Korn Ferry, April 2026

• Tobi Lutke: Shopify AI Mandate Memo — April 2025

• Gokul Rajaram: “Design Is the First AI Casualty” — LinkedIn, April 2026

• Jaya Gupta: “Experience Is Now a Tax” — X, April 2026

• Aaron Levie: Consulting Onslaught for Agent Deployment — LinkedIn, April 2026

• Todd Saunders: “The Harness Flips Vertical SaaS” — X, April 2026

• Azeem Azhar: “Everyone’s Looking for a Bubble. No One Sees the Stampede.” — Exponential View, February 2026

• Ben Lang: YC’s AI-Native Company Playbook — X, April 2026

• Nick Tran: “Expedia Just Bet Its Gen Z Strategy on a Livestreamer” — LinkedIn, April 2026

• Gallup: Rising AI Adoption Spurs Workforce Changes — April 2026

• Morgan Stanley: AI Adoption Accelerates — April 2026

My prior writing on these topics:

• Building for the Agent Experience Gap — Substack, April 2026

• The Three Adoption Playbooks — Substack, April 2026

I’ve been writing about a structural shift for months. That AI would reorganize enterprise software from the inside out. That the real challenge would be infrastructure, not generation. That connectors without architecture would be expensive and fragile.

For most of that time, these were arguments. Good ones, I think, but still arguments. The kind of thing where you lay out the logic, cite a few early signals, and wait.

Last week, I stopped waiting.

In the space of a week, seven writers working independently arrived at the same structural argument. Different industries, different audiences, no shared thread connecting them. When I lined up what they published, the overlap was hard to explain as coincidence.

Three structural shifts are now visible. Together, they’re a preview of the next eighteen months.

TL;DR

• Salesforce announced Headless 360, exposing CRM capabilities as APIs for agent consumption. The same system now serves humans through a UI and agents through APIs, adapting to whoever is consuming it. Levie’s formula: “Seats for the people, consumption for the agents.” Cursor’s $60B SpaceX/xAI deal exposed what happens when your interface layer is someone else’s model.

• PwC found 75% of AI’s economic gains captured by 20% of companies, at a 7.2x performance gap. Flanagan (former HubSpot SVP) found 98% CMO adoption but fewer than 33% see returns. His independently derived four-layer architecture echoes From Content Generation to Content OS.

• Rajaram called coding “the canary vertical,” with legal and finance 12-18 months behind. Kniberg, creator of the Spotify Model, runs his new company on four agents with daily releases. Singhal predicts 30K PMs shed, 8K rehired AI-first.

• Ramp AI Index: Anthropic gained 6.3 points in a single month, predicted to overtake OpenAI in enterprise share within 60 days. Business AI adoption crossed 50%. The strongest predictor of adoption is investor backing type.

• Gensler surveyed 16,400 workers across 16 countries. AI power users spend less time alone, more time learning and socializing. The isolation narrative doesn’t survive contact with behavioral data.

The interface went liquid

For decades, Enterprise SaaS meant a fixed interface: screens, buttons, workflows designed for a human sitting in a chair. What happened this week is that the interface started to dissolve into something malleable. The same underlying system now serves different surfaces to different consumers, adapting its shape to whoever is calling.

Salesforce made the most structurally honest move of its AI era. After years of rebranding capabilities (Einstein, Einstein GPT, Einstein Copilot, Agentforce), the company announced Headless 360: CRM capabilities exposed as APIs and tools for direct agent consumption. The human interface stays. But now there’s a parallel surface, shaped for agents, running alongside it. The same CRM, two entirely different experiences, each adapting to its consumer.

This is the shift. The interface didn’t die. It multiplied. A marketer still sees dashboards and campaign builders. An agent sees tool schemas and structured data. A workflow orchestrator sees API endpoints and permission boundaries. Each consumer gets the surface that fits how it works.

Aaron Levie has been building toward this framing in real time. In a detailed thread, he laid out the business model case: software has been constrained by how much people can do in a given day. Agents remove that constraint entirely. They work 24/7, run in parallel, and use platforms far more than humans ever did.

Instead of reviewing contracts one by one, agents review all of them. Instead of running a handful of marketing campaigns, agents run ten times more. His formula is clean: “Seats for the people, consumption for the agents.” Same platform. Two modes of engagement. The interface adapts to whoever shows up.

The same week, Cursor discovered what happens when your interface layer belongs to someone else. Saanya Ojha’s analysis of the SpaceX/xAI deal is the sharpest wrapper-vulnerability case study I’ve read. A $50B company, at peak momentum, forced to reckon with the fact that its most critical supplier is also a potential competitor. Anthropic had already demonstrated the leverage: during acquisition rumors, they cut Windsurf’s Claude access entirely. If the malleable interface is the new value layer, owning it matters. Renting it from a model lab that can revoke access is a structural risk.

Alexander Atzberger, CEO of Optimizely, declared MACH dead. The technical plumbing standards are adopted. What matters now is the orchestration layer that decides which surface serves which consumer. “The suite always wins,” he wrote, positioning Optimizely’s Opal as that orchestration platform. Whether you agree with the vendor framing or not, the structural claim is right: the static interface is commoditized. The dynamic, adaptive layer above it is where the value moved.

In March: “MCP gave us the plumbing. But most implementations forgot to build the house.” Salesforce, Levie, Cursor, and Atzberger all published evidence this week that the house, the adaptive layer that shapes itself to whoever is consuming the software, is now the only thing that matters.

The returns gap is real and widening

Here is a number that should keep every CMO awake: 98% of them use AI. Fewer than a third see the expected returns.

Kieran Flanagan, former SVP of Marketing at HubSpot, diagnosed why. The systems produce more, not better, because they lack marketing fundamentals. No audience context. No voice. No feedback loops. He cited Ries and Trout from 1981: “The mind has always been the market.” AI systems without marketing understanding produce volume without depth.

His proposed fix is where it gets interesting. A four-layer Claude Code architecture: context (audience language, voice, positioning as .md files), rules (CLAUDE.md under 60 lines), reach (live stack integration for real data), and operation (skills that compound with each run). That maps almost exactly to the four pillars I described in From Content Generation to Content Operating System: orchestration, knowledge integration, quality evaluation, and closed-loop measurement. Context is knowledge integration. Rules are orchestration. Reach is measurement. Operation is orchestration plus quality. We arrived at the same architecture from different starting points.

The pattern is visible at the enterprise level too. PwC surveyed 1,217 senior executives across 25 sectors and found that 75% of AI’s economic gains are captured by just 20% of companies. The leaders deliver 7.2x the financial performance of their peers. They are 1.9x more likely to use autonomous AI and are removing human intervention at 2.8x the rate. The gap is widening.

Stanford HAI’s data tells the same story from a different angle: 88% of organizations use AI. Single-digit percentages have mature deployments delivering real value.

And the Ramp AI Index crossed its own threshold this month: business AI adoption hit 50.4%, up from 35% a year ago. The most striking finding isn’t the adoption rate. It’s that the strongest predictor of adoption is investor backing type. VC-backed companies: 80%. PE-backed: 64%. Everyone else: 45%. VCs are functioning as transmission mechanisms, pushing tools into portfolios faster than organic adoption would.

Read those numbers together. 98% CMO adoption. 88% across all organizations. 50% business adoption. 75% of economic gains concentrated in 20% of companies. Single-digit maturity rates.

Everyone adopted. Architecture is the problem. “The generation problem is largely solved. What’s not solved is everything around it.” I wrote that in March. It was a claim. Now it’s a data point.

The canary is singing

Gokul Rajaram made the timeline concrete. In a LinkedIn post, he called coding “the canary vertical,” the first industry where AI reached mainstream disruption. The reason coding leads: objective feedback loops. Code runs or it doesn’t. Tests pass or they fail. That signal richness accelerated model capability beyond law, finance, or medicine.

His timeline: legal and finance are 12-18 months behind coding. Healthcare is further out. Every knowledge-work vertical will follow the same path, just on a delayed curve.

The same week, Henrik Kniberg, the creator of the Spotify Model, showed what the endpoint looks like. His new company, Abundly, runs on four specialized agents: one writes code, one converts Slack messages into Notion tickets, one handles releases end-to-end, and one (named Grace) coordinates the others, handles stakeholder requests, and improves herself based on team feedback. They ship platform releases daily. The engineers focus on design, architecture, and UX. Grace is “effectively improving the platform that she is running on.”

A production system at a real company, built by someone who literally wrote the book on how software teams organize.

Nikhyl Singhal extended the disruption into product management with the most aggressive headcount prediction I’ve seen: companies will shed 30,000 PMs and rehire 8,000, all AI-first. Wholesale replacement. The credentials that used to matter are devaluing fast.

At Sierra, Bret Taylor’s company scrapped its coding interview entirely. The replacement: candidates drive product ideation, build for two hours with any AI tools they choose, then demo and discuss. The evaluation criteria shifted from “can you write code” to “can you think about products and use AI with judgment.” At Intercom, the top five power users of their Claude Code deployment weren’t engineers. They were designers, PMs, and TPMs.

The canary vertical now has a concrete timeline: a production case study, a headcount prediction, and a scrapped interview format attached to it. And the pattern will replay in every knowledge-work vertical on Rajaram’s timeline.

The counter-signal

One dataset pushes against the disruption narrative.

In March, the Gensler Research Institute surveyed 16,400 office workers across 16 countries. 30% now qualify as “AI power users.” The headline finding: power users spend less time alone (37% vs. 42%), more time learning (12% vs. 8%), and more time socializing (11% vs. 9%). They report stronger team relationships and spend nearly twice as much time at client sites and coworking spaces.

The most common objection to deep AI adoption, in every change management conversation I’ve been part of, is that it isolates people. The behavioral data suggests the opposite: AI automates solitary cognitive work, freeing capacity for the collaborative work organizations actually need more of.

The disruption is real. The human cost may be more nuanced than the headlines suggest.

The Odd Find

AI’s biggest productivity boost landed somewhere nobody expected: at home.

Stanford SIEPR researchers tracked over 200,000 U.S. households and found that AI delivers 76-176% efficiency gains on digital chores: job hunting, travel planning, comparison shopping, scheduling. Dramatically, measurably faster.

What people did with the freed time is the odd part. They watched Netflix. They scrolled Instagram. The productivity gains were real, and the reinvestment went to leisure.

Younger, higher-income households adopted faster, widening the digital divide. But the finding that sticks: the productivity paradox inverted. AI makes people dramatically more efficient at the things that don’t show up in any company’s productivity metrics. The gains are real. They just landed in the wrong spreadsheet.

What I’m thinking about after this week’s reading

None of these signals are surprising in isolation. Levie has been building the headless case in public for weeks. The returns gap data has been trickling in. The coding disruption was already visible in Stanford HAI’s developer employment numbers.

What’s new is the convergence. Seven unrelated voices, in the same week, all publishing evidence that this shift is no longer theoretical. Salesforce restructuring around agents. Cursor discovering it’s a wrapper. PwC quantifying the concentration. Kniberg running a company on four agents. Flanagan independently deriving the same four-pillar content architecture.

The interface is going liquid, and the returns concentrate where the infrastructure exists. Every knowledge-work vertical now sits somewhere on Rajaram’s curve.

I should be transparent: the infrastructure layer I keep pointing to (orchestration, knowledge integration, quality evaluation, closed-loop measurement) is the problem space I work on at Typeface. But these numbers came from PwC, Stanford, Ramp, a former HubSpot SVP, and Optimizely’s CEO. The convergence is structural, not convenient.

The technology is the easier part. It always has been.

Sources:

• Saanya Ojha: “Salesforce Loses Its Head” — The Change Constant, April 2026

• Aaron Levie: “Software going headless is inevitable” — X, April 18, 2026

• Saanya Ojha: “Cursor’s Shotgun Wedding” — The Change Constant, April 2026

• Alexander Atzberger: “The New MarTech Stack” — LinkedIn, April 2026

• PwC: 2026 AI Performance Study — April 2026

• Kieran Flanagan: “The Problem with AI Marketing Systems” — LinkedIn, April 2026

• Stanford HAI: AI Index Report 2026 — April 2026

• Ramp AI Index: April 2026 — April 2026

• Gokul Rajaram: “Coding: The Canary Vertical” — LinkedIn, April 2026

• Henrik Kniberg: “When Coding Is No Longer the Bottleneck” — LinkedIn, April 2026

• Nikhyl Singhal: “Why Half of Product Managers Are in Trouble” — Lenny’s Newsletter, April 2026

• Sierra: “The AI-Native Interview” — April 2026

• Intercom: How Intercom 2x’d Engineering Velocity — Lenny’s Newsletter, April 2026

• Gensler: 2026 Global Workplace Survey — March 2026

• Stanford SIEPR: AI Productivity Boost on Digital Chores — April 2026

My prior writing on these topics:

• To MCP or Not to MCP — LinkedIn, March 2026

• From Content Generation to Content Operating System — Substack, March 2026

Nobody cited anybody else. Each arrived at the same profile independently: someone who can map a business process end to end, identify where an agent replaces a bottleneck, build or configure the system, and manage it in production. Technical enough to wire up MCP (Model Context Protocol) servers and CLI tools. Process-fluent enough to know which workflows matter. Politically skilled enough to get buy-in.

That convergence was the most interesting signal of the week. But so was the gap between these emerging job descriptions and what the hiring data actually says.

TL;DR

• Three voices, three platforms, one job: Levie described the “agent deployer,” Chow described the “marketing engineer,” and Mehra described the “AI conductor.” None cited each other. All described someone who maps business processes and wires up agent systems.

• Korn Ferry surveyed talent acquisition leaders and the #1 skill for 2026 is critical thinking. AI skills rank fifth. The meta-skill beats the tool skill.

• Garry Tan published the most specific agent architecture breakdown I’ve seen: five definitions, a three-layer stack, and a feedback loop tested at YC Startup School with 6,000 founders.

• Anthropic’s run rate surged from $14B to $30B in two months. A leaked OpenAI memo positions them as a “single-product company.” Both companies plan to IPO this year.

• Allbirds sold its sneaker brand for $39M, rebranded as NewBird AI, and is now selling GPU-as-a-Service. The stock popped.

The role that appeared in three inboxes

Levie’s version is the broadest. From his enterprise conversations, every functional team will need its own agent deployer. Distributed practitioners, reporting to IT or AI leadership, with full autonomy to connect systems and drive automation. His focus is 100x workflows: lead processing, contract review, client onboarding. Wholesale replacement of entire workflow steps.

He followed up later in the week by arguing that forward deployed engineers are more critical in the agent era. Selling agents, he wrote, “is far closer to a customer buying from a professional services firm than implementing traditional technology.” The four requirements: deep domain understanding, system wiring, context setup, and change management. Every system integrator and consulting firm is spinning up new practice areas around this.

Chow’s version is narrower and equally specific. At Google, the marketing engineer orchestrates AI systems to solve real problems, “rather than just typing prompts into a box.” His examples are pipelines: brand mention sentiment alerting, real-time competitor battle card updates, rapid prototyping. The hiring framework starts with process mapping, moves to bottleneck identification, then building, then measuring pipeline impact. “Build it” is step three.

Mehra’s version comes from the C-suite. The future CMO is a “five-headed hydra”: AI Conductor, Financial Strategist, Community Architect, Cultural Decoder, Talent Multiplier. The AI Conductor label is the cleanest articulation I’ve seen of what fluency looks like when it becomes a dedicated leadership competency.

Three angles, one archetype.

And then there’s data that complicates the picture. Korn Ferry published their 2026 talent acquisition report in early April, and it landed in my scan this week.

Their 2026 talent acquisition report surveyed TA leaders and found that 73% rank critical thinking as the #1 skill for the year. AI skills ranked fifth. The meta-skill beats the tool skill. Employers want people who can assess AI output, spot flaws, and know when to override. They also found that 52% plan to add autonomous AI agents as team members, and only 22% believe leaders can effectively manage human-AI teams despite billions in investment.

The tension is productive: organizations are naming agent-era roles that require process mapping, domain expertise, and change management. And the scarcest capability isn’t AI proficiency. It’s judgment.

The architecture that got specific enough to argue about

While the job titles were crystallizing, the architecture underneath them crossed from theory to playbook.

Garry Tan published the most precise breakdown I’ve read. The YC president and CEO laid out five definitions that separate 2x productivity from 100x:

• Skill files: reusable procedures the model follows.

• The harness: thin, just runs the model in a loop and manages context.

• Resolvers: routing tables that load the right context for each task.

• The latent/deterministic split: the most common agent design mistake is forcing deterministic problems into LLM territory.

• Diarization: structured profiles synthesized from dozens of documents, where the model holds contradictions and makes judgment calls.

His CLAUDE.md was 20,000 lines before he realized it was degrading model attention. The fix: roughly 200 lines of pointers.

The three-layer stack: fat skills on top (90% of the value), thin CLI harness in the middle, deterministic tooling on the bottom. Push intelligence up, push execution down, keep the middle thin. At YC Startup School, with 6,000 founders, a self-improving feedback loop brought “OK” ratings from 12% down to 4%.

Adam Miller made the leap from architecture to product. He wrote about Goose, an open-source agent runtime under the Agentic AI Foundation (a Linux Foundation umbrella body), and his central claim is blunt: over 90% of agent use cases don’t require custom development. “Stop building agents, start harnessing them.” Goose’s three-element design mirrors Tan’s stack: interface, agent runtime, pluggable MCP extensions. The competitive layer, Miller argues, is “Harness Engineering,” building the enablement layer around agents rather than reimplementing core functionality.

Andrew Ng formalized the same pattern for individual developers. Spec-driven development: write a specification that defines what to build, then work with a coding agent to implement it. The spec is persistent context that survives across sessions. Ng’s claim is that many of the best developers already work this way.

Ethan Mollick, writing in late March, added the interface layer. The gap between AI capability and actual usefulness, he argued, is an interface problem. Most users interact through chatbots that dump walls of text. Claude Dispatch decouples the control interface from the execution environment: you direct the agent from your phone while it works on desktop files. Adaptive interfaces, where the AI generates the right tool for each moment, are where Mollick sees this heading. If he’s right, the harness needs an interface layer on top of Tan’s three-layer stack.

From the archives: Maria Weaver published something back in January that I only found this week, and it reframes everything above. She applied the skills pattern to personal life. She built a chained Claude Code system for weekly, monthly, and quarterly planning. Each skill reads the output of the previous layer. Calendar reminders trigger the cadence. Her framing: Claude Code skills are externalized “implementation intentions” from behavioral psychology, if-then plans that work through automaticity. For her ADHD brain, the skill file holds the behavioral link that unreliable memory can’t. “The system doesn’t make me more disciplined. It makes discipline irrelevant.”

The numbers that explain why every vendor sounds the same

The architecture is getting specific. The money following it is getting specific too.

Anthropic’s run rate surged from $14B to $30B between February and April. Enterprise contracts exceeding $1M are routine. An IPO is planned within six months.

Meanwhile, a leaked four-page memo from OpenAI’s chief revenue officer laid out a different strategy. Five enterprise priorities: own the model layer, own the agent platform, expand through Amazon distribution, sell the integrated stack, own deployment. The language toward Anthropic was aggressive, framing them as ideologically restrictive and alleging an $8B run-rate inflation through accounting treatment. The memo’s most revealing line: “the companies that win enterprise AI will not just have the best models. They will have the best ability to get those models deployed.”

Saanya Ojha read the same signals differently. OpenAI, she argued, is executing a quiet pivot from spectacle to “workflow position.” GPT-Rosalind (a biology-focused reasoning model) signals vertical specialization. An expanded Codex, now with workflow automation and memory, signals horizontal platform play. While Anthropic dominates headlines, OpenAI is building switching costs through infrastructure embeddedness.

The question neither company is answering directly: why does every vendor pitch sound the same? Jaya Gupta provided the structural explanation. Before AI, enterprise data had clear layers with clear owners. CRM data had a CRM buyer. Warehouse data had a data team. Now, state spans five or more layers simultaneously with no consolidation, unclear ownership, and every layer interacting with every other. Every AI vendor genuinely needs access to the same data. So every vendor makes the same claim: “We’re the safe, trustworthy, enterprise-grade way to give AI access to your data.” The convergence is structural.

CIOs are being pitched fifty times a week on essentially the same story by companies doing genuinely different things. Gupta’s diagnostic for enterprises that want to cut through: “Which of the state layers do you own, how does it interact with the others, and what happens to the state you accumulate when we end the relationship?”

The Odd Find

A sneaker company walked into a GPU.

Allbirds, once valued at roughly $4B, sold its footwear brand for $39M, raised $50M in convertible financing, rebranded as NewBird AI, and announced a GPU-as-a-Service business. The stock surged. Saanya Ojha placed it in a lineage: Long Island Iced Tea rebranding as Long Blockchain in 2017 (stock tripled overnight), Kodak’s crypto pivot in 2018 (stock tripled), Bioptix becoming Riot Blockchain. When a company’s original business collapses, the ticker symbol itself becomes the primary asset. A publicly traded vehicle for whatever narrative the market currently rewards.

What I’m thinking about after this week’s reading

Azeem Azhar shared a Nature Communications paper this week that reframed everything else I read. Researchers studied 285 polities across six continents over 10,000 years and found two quantified thresholds: a scale threshold and an information-processing threshold. Civilizations grow until they hit a coordination ceiling. The ones that survive invent new tools for processing information: writing, currency, bureaucracy. The ones that don’t, collapse.

Azhar’s framing: we’re at one of those ceilings now.

Mike Fisher, writing about product organizations, made a version of the same argument at a smaller scale. The fundamental failure mode is confusing outputs with outcomes. He told the Langley vs. Wright brothers story: Langley optimized for visible progress and failed. The Wrights optimized for understanding the problem and succeeded nine days later. ROI, Fisher argues, is “a terrible product manager” because it cannot see second-order effects like trust, habit formation, or emotional resonance. Good teams measure learning velocity, not feature count.

The thread I keep pulling: the new roles being named right now (agent deployer, marketing engineer, AI conductor) are information-processing innovations. They exist because organizations hit a coordination ceiling they can’t clear with existing structures. The architecture getting specific (thin harnesses, fat skills, spec-driven development) is the tooling response. The platform war ($30B run rates, deployment strategy, vendor convergence) is the market’s attempt to own the infrastructure layer beneath it all.

And the Korn Ferry finding that critical thinking outranks AI skills is the quiet signal underneath the noise. The tools change. The judgment doesn’t. That’s what the new roles actually require.

Sources

• Aaron Levie on the agent deployer role

• Aaron Levie on forward deployed engineers

• Marvin Chow on marketing engineers

• Reggie Panaligan on Vineet Mehra’s five-headed hydra CMO

• Korn Ferry: TA Trends 2026

• Garry Tan on thin harness, fat skills

• Adam Miller on Goose and harness engineering

• Andrew Ng on spec-driven development

• Ethan Mollick on Claude Dispatch and interfaces

• Maria Weaver on Claude Code for personal planning

• Michael Spencer on Anthropic as biggest AIaaS

• OpenAI enterprise memo via The Verge

• Saanya Ojha on OpenAI’s pivot

• Jaya Gupta on enterprise AI state explosion

• Saanya Ojha on Allbirds/NewBird AI

• Azeem Azhar on information-processing ceilings

• Mike Fisher on building the right thing

Andrej Karpathy posted about maintaining a personal wiki with an LLM. A VC described the AI chief of staff he built for his practice. Within 48 hours of Karpathy’s post, someone shipped an open-source version. By the end of the week, a top AI advisor had built one with a feature that challenges your own assumptions.

Some of these builders explicitly cited Karpathy. Others arrived at the same shape independently, solving different problems in different domains. That’s what caught my attention: the convergence wasn’t coordinated. Raw sources go in. A structured wiki comes out. The human writes the rules. The LLM does the rest.

Here’s what I took away from the week’s reading.

The piece that reframed how I think about search

Karpathy’s thread got 1.2 million views, and I think it’s because he named something practitioners already felt. If you care about a body of knowledge, retrieval-augmented generation is the wrong tool. You don’t want to embed documents and hope the right chunks surface when you ask a question. You want to process them upfront: summaries, entity pages, cross-references, backlinks. A compiled wiki. The search index was the old answer.

The distinction sounds academic until you see the numbers. Graphify, the first open-source implementation, claims 71.5x fewer tokens per query compared to reading raw files. That’s the difference between a system that scales and one that bleeds money.

What I found more interesting than the pattern itself was how fast people adapted it to wildly different domains.

Ryan Sarver, a VC, built “Stella” as his AI chief of staff. He wasn’t referencing Karpathy. He was solving his own problem: meeting prep, relationship tracking, task management. But the architecture he landed on is structurally identical. Two-layer memory: raw daily notes, then a curated file synthesized from those notes. His design rule stuck with me: LLMs handle judgment, scripts handle everything deterministic. No mixing.

The ones that explicitly built on Karpathy’s post moved fast. Allie K. Miller built Claudeopedia over a weekend and added a feature I hadn’t seen anywhere else: an automated job that reads your recent writing against the wiki and challenges your own assumptions. A reflexive layer that asks whether you’re repeating yourself or missing something.

None of these are coding tools. They’re personal operating systems running on a coding tool’s infrastructure.

The reading that convinced me infrastructure matters more than models

Akshay Pachaar published what might be the most thorough breakdown of agent harness architecture I’ve read: 12 components, from orchestration loops to verification systems to subagent coordination. The headline finding was stark. Changing only the infrastructure, keeping the same model, moved agents 20+ ranking positions on benchmarks.

Nicholas Charriere took this a step further with his “Great Convergence” thesis. By the end of 2026, he argues, app companies, model companies, and infrastructure companies will all look like they’re building the same product. The general agent harness is the shared surface. Claude Code made it popular for coding. Now it works for anything you can do on a computer.

I kept thinking about that while reading about Hannah Stulberg’s work at DoorDash. She’s a former Google APM who scaled Claude Code into what she calls a “Team OS” supporting 20+ people across product, analytics, and operations. The system has seven components: nested configuration files, shared skills, an analytics layer, and a launch gate that won’t let you ship a feature until all operational knowledge is compiled into the repo. That last detail is the one I keep coming back to. Compiled knowledge as a prerequisite for launching. The afterthought era is over.

I wrote last week about building a customer dashboard over a weekend that changed how 200 people think about accounts. Aman Khan made the broader case: domain expertise packaged as skills is becoming a go-to-market channel. Companies will ship skills alongside products. The barrier drops from a months-long engineering project to cloning a repository.

Charriere also released Meta-Harness, a method for autonomously optimizing these harness configurations end-to-end. If infrastructure quality outweighs model quality for agent performance, and the benchmarks suggest it does, then automating infrastructure optimization is a compounding advantage.

All of which raises a question: if the infrastructure is shifting this fast, what does that mean for the people expected to operate it?

The hiring bar that just went public

Zapier published the second version of its AI fluency rubric, and it verbalizes what we’ve been observing and working towards in our own hiring at Typeface. The first version asked whether you could use AI tools. The second version asks whether you build repeatable systems with them, whether you’re improving, and whether you catch errors before they ship.

Four dimensions: mindset, strategy, building, accountability. Accountability is the new addition, and it’s pointed. “With AI, you can delegate the work, but not the accountability.” Zapier now observes candidates using AI in real time during skills assessments, watching how they iterate rather than whether they produce a polished first output.

Aakash Gupta documented what this looks like in practice for PM roles specifically: Google, Figma, and Perplexity now run 45-minute “vibe coding” rounds where candidates build working prototypes with tools like Cursor. The rubric isn’t theoretical. It’s already in interview loops.

The most consequential change is from snapshot to slope. V2 measures the trajectory of someone’s fluency, not where they are today. Hiring for growth rate. That’s a significant philosophical shift for any company that takes capability assessment seriously.

Adriane Schwager‘s reaction added necessary context. The floor has moved in tech, but it hasn’t moved everywhere. A workflow Zapier now rejects as insufficient, using AI for first drafts and editing manually, would be considered advanced at most accounting firms and small agencies. She frames the gap as an opportunity, not a failure. I think she’s right, and I think most people building AI hiring rubrics aren’t thinking about this asymmetry.

Austin Lau at Anthropic offered a complementary lens: four ways growth teams use AI. Automate existing work, use AI as a thought partner, tackle work that used to fall below the ROI threshold, and build bespoke tools. His argument is that the third dimension is the most underused and the real compounding unlock. Practitioners stuck in the first two have plateaued.

What I’m thinking about after this week’s reading

Jack Dorsey published a piece arguing that AI eliminates the need for management hierarchy. The org chart becomes an interface. Jaya Gupta made a related argument about context graphs: whoever accumulates structured decision traces first owns future thinking. The enterprise moat is the history of your judgment calls, not your current headcount.

Shreyas Doshi made the individual version of the same point: product sense, the ability to improve on what AI already produces, is the only durable career moat as tools get commoditized.

The thread connecting all of this week’s reading is that the middle layer is settling. Between the models below and the applications above, an infrastructure layer is taking shape. Karpathy described it for knowledge. Pachaar mapped it for agents. Charriere says the two are converging. DoorDash is running a cross-functional team on it. Zapier is hiring for the fluency to operate within it.

The question I keep sitting with: the infrastructure is arriving faster than most organizations can absorb it. Some teams are already running on compiled knowledge and shared agent harnesses. Others are still debating whether to allow ChatGPT. The distance between those two groups grew wider this week, and nothing I read suggested it’s going to close.

Before the closing, a scenic detour.

The Odd Find

This one’s all good taste, no AI.

Jolyon Varley resurfaced a Coca-Cola campaign from 2024 that somehow flew past me. Someone at Coke looked at the 🥤 emoji sitting on every phone in the world and thought: that’s ours. What followed was Emoji Coke Cups in Saudi Arabia, a gamified experience pairing food emojis to unlock restaurant vouchers, 630 million impressions, 62% conversion rate. The entire strategy fits on a napkin: pay attention to what’s already part of culture. Thank you, Jolyon, for the time travel.

On day three, something shifted. I’d been building my own version of the Karpathy pattern, a knowledge bot called Yara (”friend” in Hindi). Forty sources in, structured wiki out, running on Claude Code. I asked it a question that spanned sources I’d ingested days apart, and it answered with connections I’d forgotten I’d read. That’s when the pattern stopped being theoretical.

Building the thing taught me more than reading about it ever could. The schema matters in ways the blog posts don’t cover. The compilation step changes how you think about what’s worth ingesting. The linting catches gaps you didn’t know you had. And this newsletter was drafted from the wiki Yara maintains.

Sources

• Andrej Karpathy on LLM knowledge bases

• Graphify: open-source knowledge graph

• Ryan Sarver on building an AI chief of staff

• Allie K. Miller on Claudeopedia

• Akshay Pachaar on agent harness architecture

• Nicholas Charriere on the Great Convergence and Meta-Harness

• Aakash Gupta on Team OS with Claude Code

• Aman Khan: Every PM Should Be Building Skills

• Zapier: Raising the AI Fluency Bar in Hiring

• Adriane Schwager on Zapier’s hiring rubric

• Austin Lau on AI for growth teams

• Block: From Hierarchy to Intelligence

• Jaya Gupta on context graphs

• Shreyas Doshi on product sense as career moat

• Aakash Gupta: AI PM Interview Guide 2026

• Jolyon Varley on Coca-Cola’s emoji campaign

• The Drum: Coca-Cola claims the soda emoji