The 5 Ocean Beach bus pulled up to the stop on Market Street wrapped in black vinyl, one word repeating in white block letters down its entire length: AGENTS. AGENTS. AGENTS. I stood on the brick sidewalk outside Moscone West, badge around my neck, watching a city bus advertise the exact thing I had flown across the continent to spend three days thinking about.

That bus is the most honest picture of San Francisco right now. Inside the conference: 7,000 engineers comparing notes on what actually works. Outside: a city where even the public transit has an agent strategy, while the rest of the economy, the one you and I work in, is still trying to get one pilot past legal.

This issue is everything I brought back for you. Not “look where I went.” More like: here is what 7,000 of the people building this stuff are converging on, why it matters for your roadmap, and what I would do about it starting Monday. Let’s go.

The setting: high signal, low BS, one bubble

First, the vibe check, because context matters for everything that follows.

AI Engineer World’s Fair is the biggest gathering of AI engineers in the world: 18 tracks of keynotes, breakouts, and workshops. It was extremely well organized, well attended, and refreshingly free of vaporware. Talks came with architecture diagrams instead of adoption curves. Demos ran live. When something failed on stage, speakers said so. My kind of conference.

The swag table handed out prescription bottles labeled AGI PILLS (”Intelligence, Compressed”). For a brief moment I thought of the Matrix and hoped for blue/red pills. Filled with candy, sadly. I checked.

And that is the tension you need to hold while reading everything below. This was San Francisco talking to San Francisco. The bus wraps, the AGI pills, the people walking around with laptops open so their agents can keep working (a real thing Addy Osmani teased from the main stage): none of that reflects the real economy yet. Enterprise adoption where you live is still messy, budget-constrained, and skeptical.

But here is what made this year different, and it is the reason I am writing this issue: it was not just builders. I met a noticeable number of engineering leaders, VPs and directors and staff-plus folks, walking the same halls trying to make sense of it all for their organizations. The questions I heard in hallway conversations were not “which model is best.” They were “how do I run this safely,” “what do I stop building,” and “who owns the output.” Those are translation questions. That is exactly the gap this newsletter lives in.

As a bonus, I ran into an old friend from school and went to a couple of sessions together. He’s also deep-building the future but also learning what others are doing to trade notes. That’s what I loved the most about this environment.

Five signals stood out. Each one comes with the “so what” for your enterprise.

Signal 1: The model race quietly became the harness race

Nobody on any stage argued their model was the moat. Read that again. At the world’s largest AI engineering conference, in the most model-obsessed city on Earth, the models were treated as interchangeable engines. All the energy was one layer up: the harness, the loop, the system around the model.

OpenAI’s Codex keynote spent its entire second half walking through their stack as a set of open layers anyone can build on: the API, the open-source harness, the app server, the plugins, even AGENTS.md as a deliberately generic file format other tools can adopt. Their words: “we’re not building one system for OpenAI and a second system for developers.” The model is the engine; everything they showed was chassis.

swyx (the conference organizer) captured it in one diagram he called Loopcraft: six nested loops, from the token loop at the bottom, up through chat, agent turns, goal loops, automations, and at the top, software factories. Each level wraps the one below it with judgment about when to exit. The engineering has moved from “prompt the model” to “design the loops.”

If you have been reading this newsletter, this should sound familiar. Issue #20 (”The Model Doesn’t Matter Anymore”) made exactly this argument: the frontier converged, and the harness is the part you actually own. Three days in San Francisco turned that thesis from my opinion into consensus.

What this means for you: stop budgeting for model selection like it is a strategic decision. It is a procurement decision now. The strategic investment is the harness layer: context, tools, memory, and the loops that decide when work is done. That is where your differentiation lives, and unlike the model, it compounds.

Signal 2: Intelligence is getting cheaper per token and more expensive per task

This is the most useful contradiction in enterprise AI right now, and the closing session (from the benchmarking team at Artificial Analysis) put real numbers on it.

Token prices keep collapsing: 5 to 10x cheaper every year for the same level of intelligence. OpenAI announced GPT-5.6 Sol at $1 per million input tokens and $6 per million output, delivering roughly their previous flagship’s intelligence. Then they showed it running on Cerebras hardware at 750 tokens per second.

And yet: everyone’s AI bill is going up. Artificial Analysis showed why. As models get more capable, we hand them longer, harder, more agentic tasks. On their new AA-Briefcase benchmark (realistic multi-week knowledge work: data science, banking operations, strategy), single tasks now regularly cost over $20 in tokens. Cheap tokens, expensive tasks.

What this means for you: cost per token is the wrong unit for planning, negotiating, or reporting. The unit that matters is cost per completed task. Two models with identical per-token pricing can differ by multiples on what a finished piece of work costs, because one takes three attempts and burns tokens thinking. If your finance team is modeling AI spend on token prices, they are modeling the wrong thing.

I am not leaving this one as a hot take. Next week’s Build Log puts numbers behind it: same tasks, several models, measured cost per completed task. Consider this the field report; next week is the lab report.

Signal 3: Your new job title is Verdict Owner

The best talk of the conference was Addy Osmani’s closing keynote on keeping humans in the loop, and it gave language to something every engineering leader I spoke with was circling.

His frame:

• Quality is the system of checks that produces evidence.

• Verdict is the human, accountable decision made from that evidence.

• Answerability is the ability to explain and stand behind that verdict later. Agents can produce the first. Only people can own the second and third.

He backed it with numbers that should worry you more than any layoff headline: in one study he cited, when the AI was wrong, 73% of people went with the wrong answer anyway, feeling more confident. And while nearly everyone says they distrust AI-generated code, only about half consistently verify it before committing. Distrust without bandwidth. The bottleneck is not generation anymore; it is verification.

His operational rule was the most quotable line of the week: explain it or don’t ship it. Not because a human types every line, but because someone must understand the work well enough to defend it in production.

If you’re making the case for AI review capacity internally, here’s the framing that works: “Generation now scales faster than comprehension, so fund the verification layer like you funded the build layer, or accept that nobody in the building can explain what we ship.”

What this means for you: start assigning verdict ownership explicitly, the way large codebases assign OWNERS files. For every agent workflow in your company, one name should answer three questions: what evidence does this produce, who decides ship or block, and who explains it when it breaks. If the answer is “the team,” you do not have an owner. You have a future incident review.

Signal 4: The enterprise plumbing finally showed up

Two years ago this conference was demos. This year, the breakout rooms were full of the unglamorous stuff that actually gates enterprise deployment: knowledge, governance, and isolation.

Microsoft’s Pablo Castro gave a genuinely useful taxonomy of the knowledge an agent needs: intrinsic (what is in the model), extrinsic (what you ground it on: documents, warehouses, the web), and learned (what accumulates from doing the work: the feedback loop that captures how your organization operates). Two details stood out. Every knowledge base in their Foundry platform is exposed as an MCP server, so any harness can plug in without glue code. And their Agent Optimizer closes the learned-knowledge loop automatically: evaluate a baseline, generate candidate configurations, hill-climb the eval, apply the winner. Agents that tune agents, in a product, today.

The security breakouts were even more telling. A few slides from my camera roll:

“Architecture decides the blast radius.” “A useful coding agent is a supply chain actor, whether you planned for that or not.” “Deterministic where we can, agentic where we must.” Entire talks were dedicated to running agents inside microVMs with policy-enforced egress. This is what a technology looks like when it stops being a demo and starts being infrastructure.

What this means for you: the vendors have moved past “can it work” to “can it be governed.” So your evaluation criteria should too. When you assess any agent platform this quarter, ask three questions: what is the isolation boundary, how does knowledge get in (and stay fresh), and where does the learning loop live. Any vendor without crisp answers is selling you a demo.

Signal 5: Open-weight is a strategy, not a discount

The talk of the expo floor was GLM-Agentic 2, the new open-weight model from Zhipu AI (delivered by remote keynote from Singapore, since the speaker could not get into the US, which is its own signal about where this field is being built). Benchmarks close to frontier on coding and agentic tasks, downloadable from Hugging Face.

More interesting than the numbers was the why. Zhipu’s reasoning for open-weighting: enterprises and governments that need security and control get on-prem deployment; companies that need domain depth get to fine-tune (Harvey, the legal AI company, fine-tunes GLM); and partners who want to co-design the future get to see the architecture. That is a coherent enterprise strategy, not charity.

Artificial Analysis put the trend line under it: for three years running, open-weight models have trailed the absolute frontier by a remarkably consistent 3 to 9 months. Their on-stage prediction: within nine months of any frontier release, someone gives away a model that smart.

What this means for you: if your data cannot leave your walls (regulated industry, sovereignty requirements, or just paranoid legal, no judgment), your penalty for staying open-weight and on-prem is now measured in months of capability lag, not years. That changes the build-vs-buy math for a lot of European and regulated-industry readers on this list. Model your architecture on the assumption that today’s frontier capability is on-prem-available within three quarters.

The part I keep thinking about

One more thing, because it reframed the whole week for me. Kwindla Kramer (Daily) gave a talk placing this moment in 80 years of computing history, and his analogy was surgical: agents in 2026 are web pages in 1995. The thing everyone is obsessed with, the thing on the side of the bus, is real and durable. But it is also the primitive, not the destination. We talked about web pages for a decade; then web pages became the boring substrate of things we actually cared about.

So when the agent hype exhausts you (it will), remember 1995. The correct response to “web pages are overhyped” was not to ignore the web. It was to learn HTML and then think hard about what comes after the page. Same move now: learn the loops, then think about what your organization builds once agents are boring.

Every signal in this issue points to the same place. If the model is interchangeable, if cost depends on task completion, not tokens, if humans must own the verdict, and if governance requires knowing what your agents actually did, then observability is the connective tissue underneath all of it. You cannot assign a verdict owner without evidence. You cannot measure cost per completed task without traces. You cannot govern what you cannot see.

The conversations I kept having on the expo floor confirmed it. Teams are no longer asking “should we instrument our agents?” They are asking “how fast can we get traces into production?” That shift, from whether to how fast, is the most reliable signal that a technology has crossed from optional to infrastructure.

What you should do Monday

The whole conference, compressed into five moves:

1. Rebalance one budget line. Take whatever you were about to spend on model evaluation and move half of it to harness engineering: context, tools, evals, memory. (Signal 1)

2. Change one metric. Add cost-per-completed-task to whatever AI spend dashboard you have. Even a rough version will change decisions. (Signal 2)

3. Name one verdict owner. Pick your highest-stakes agent workflow and put a single name on ship/block/redirect. Write it down where the team can see it. (Signal 3)

4. Add three questions to your vendor template. Isolation boundary, knowledge freshness, learning loop. (Signal 4)

5. Run the open-weight math once. If data residency matters to you, price an on-prem open-weight deployment against your current API spend, assuming a 6-month capability lag. The answer may surprise you. (Signal 5)

The frontier labs spent the week competing on who serves intelligence cheapest. The engineers spent it deciding what deserves to be built with it.

Which of these five signals is already showing up inside your company? Hit reply and tell me, I read every one.

Share this with the CTO or VP Engineering you think would find these patterns familiar. And if this pattern sounds like your quarter, you are not alone: I work with teams navigating exactly this. Reply and let’s compare notes.

Share

Know someone building with AI or navigating AI transformation? Forward this issue. It takes 5 seconds and it is the single best way to help this newsletter grow.

Signal Stack

Seven headlines from the conference floor, with my take:

1. ↑ OpenAI ships GPT-5.6 series preview; Sol tier at $1/M input, $6/M output. Frontier-class intelligence at commodity prices. If you locked pricing with any AI vendor more than two quarters ago, renegotiate. OpenAI announcement

2. 👀 Claude is now generally available in Microsoft Foundry. Announced on stage day one. Model distribution now runs through cloud catalogs; multi-model stacks are the enterprise default, not the exception. Anthropic blog

3. 👀 Zhipu AI premieres GLM-Agentic 2, open-weight, near-frontier on agentic benchmarks. Harvey already fine-tunes its predecessor. Open-weight is now a sovereignty and differentiation strategy with a 3-9 month lag, not a compromise. [DON: URL needed, see note below]

4. ↑ Artificial Analysis launches AA-Briefcase, an agentic knowledge-work benchmark. Multi-week professional scenarios, graded on rubric, analysis, and presentation. Benchmarks are finally measuring work instead of trivia. Watch the $20+ per-task costs. Artificial Analysis

5. 👀 Tavus rebuilt Apple’s 1987 Knowledge Navigator concept video with real, shipping technology. One take, four minutes, no compositing. Thirty-nine years from concept to product is either depressing or thrilling; I choose thrilling. [DON: URL needed, see note below]

6. ↑ First-ever AI Engineer Kids Day. Kids building 3D games with coding agents at the world’s biggest AI engineering conference. The talent funnel for 2040 opened this week. AI Engineer World’s Fair

Less Noise. More Signal.