Watch on YouTube · subscribe for the daily brief

Agents inherit your documentation debt in full. If your metrics are undefined, the agent will be fluently wrong at scale.

This week your AI agent got the keys to the house. Google said it will let any AI agent control your Google Home devices. Anthropic folded its Cowork agent into Claude and started building Docs and Slides to go after Google directly. Perplexity's agents helped build a production database, though someone wisely did not let them run it. Agents are everywhere now, with their hands on real systems.

And the most useful announcement of the week for people who build had nothing to do with any of that. It came from a data conference in Las Vegas, and the message was blunt: your AI agents need you to write everything down.

That line is the title of a Starburst post, and it is also, under the branding, the entire theme of dbt Summit 2026. Fivetran and dbt Labs, now merged, shipped dbt v2, a full Rust rewrite that parses a 10,000-model project up to 10x faster and, more to the point, gives an agent real-time feedback on errors, column checks, and lineage before it runs a single line. They shipped dbt State, which lets teams skip unchanged models. And they debuted something called, with no apparent irony, a Context Layer. The framing in their own announcement is the tell: as enterprises put agents into production, they need trusted data and context the business controls.

Sit with that, because it is the quiet admission under a year of agent hype. An agent is only as good as the machine-readable context it can read. Not the model. The context. And building that context, defining your metrics, describing your tables, writing down what each model means and what a correct answer looks like, is the unglamorous data-engineering work that was always the real job.

Regular readers know this is the oldest thread here. The semantic layer became the most valuable work in data engineering the moment AI removed the human who used to absorb the ambiguity. A person reading a vague ticket fills the gaps from context they carry in their head. An agent does not. It fills them confidently and wrongly. So the gaps have to move out of people's heads and into something a machine can read. dbt Summit 2026 is that idea turning into a product line, complete with a component named for it.

It also finishes a point this week kept circling. Yesterday the lesson was to put deterministic code between the model and any consequence, so a bad output cannot book the flight or run the database. Today is the input side of the same discipline: put written-down context between the model and your data, so it has something true to reason from. Bounded execution on one end, governed context on the other. That is what a production agent actually is. The intelligence is the cheap part in the middle.

Say the skeptical part too. Agentic AI is still sold as plug-and-play, and dbt Summit is the industry quietly conceding that agents inherit your documentation debt in full. If your metrics are undefined and your tables undescribed, the agent will be fluently wrong at scale. A Context Layer, and an agent that has, in the vendor's words, full context on your codebase, are shovels being sold for a hole you still have to dig. And when the pitch is an open, vendor-neutral lakehouse from a company that just doubled in size by merging, keep one eyebrow up.

Meanwhile the safety story matured from slogans into an auditing fight. Anthropic and OpenAI now both back embedded third-party evaluators. The evaluators promptly answered that embedded is not the same as independent, which is the whole point. Anthropic's own policy chief told CNBC that safety cannot rely on companies policing themselves, and critics argued the labs should shut the front door first, meaning fix the containment that let agents escape their sandboxes before hiring auditors to watch them. The EU president warned that agents escaping their environment are a preview of what is coming. A new job is forming from all of it, the developer as employee-like evaluator, and a former OpenAI researcher shipped a model built to judge options rather than generate text. The evaluator is becoming both a role and a model type. But access you grant and can revoke is not independence, and that knot is still tied.

Builder's bench

• The AI buildout has created a memory shortage severe enough that SK Hynix is in talks to make chips in the US for the first time, possibly in Intel's stalled Ohio fab. Reported and exploratory, but the shortage is real and it is in your cloud bill.
• Apple is reportedly building its own AI servers with M-series Ultra chips, reversing its exit from the server business to control its own inference. Watch whether owning the silicon becomes a moat.
• Damien Benveniste has a clear piece on how sparse attention makes long-context models cheaper and what it quietly gives up. Read it before you trust a million-token window.
• Google DeepMind spun up an institute for the big AGI questions the same week nearly 1 in 5 AI researchers were reported to have expected an extinction scenario back in 2024. File both under the discourse.
• In the slop column: a 2.5-hour AI-generated Odyssey drew reviews calling it 2.5 hours too long, and AI dating-app scams are now a genre. The floor is filling faster than the ceiling is rising.

Through-line: the agents got hands this week. What decides whether they are useful is not the model. It is the context you wrote down, and that is still your job.

One thing to do today: before you point an agent at your data, write the context it needs in a form a machine can read: metric and semantic definitions, column descriptions, what each model means, and what a correct answer looks like. The agent does not reduce that work. It raises the return on it. Do the boring modeling first. Right now it is the highest-leverage AI work on your desk.

Sources covered

1 story · 60 sources (49 news, 5 company, 6 voices)

1. datasette 1.0a40
   🎙️Simon Willison↗🎙️Simon Willison↗📰TechCrunch↗📰SiliconANGLE↗📰The Verge↗📰TechCrunch↗🎙️Simon Willison↗📰Ars Technica↗📰The New Stack↗📰TechCrunch↗📰The Verge↗📰The Verge↗📰TechCrunch↗🎙️Simon Willison↗📰SiliconANGLE↗📰CNBC↗📰dbt Labs Blog↗📰dbt Labs Blog↗📰SiliconANGLE↗📰The Decoder↗🎙️Damien Benveniste↗📰TechCrunch↗📰The Decoder↗📰SiliconANGLE↗📰BleepingComputer↗📰The Verge↗🎙️Ben Lorica↗📰The Decoder↗📰The Verge↗📰TechCrunch↗📰The New Stack↗📰The Hill↗📰The Decoder↗📰The Verge↗📰TechCrunch↗📰The New Stack↗📰The Decoder↗📰The Decoder↗📰TechCrunch↗📰The Decoder↗📰The Verge↗📰InfoQ↗📰TechCrunch↗📰TechCrunch↗📰TechCrunch↗📰The Hackers News↗📰TechCrunch↗📰MIT Technology Review↗📰BleepingComputer↗🏢Starburst↗📰The Verge↗📰InfoQ↗📰InfoQ↗📰The Decoder↗🏢NVIDIA↗📰Hacker News↗📰InfoQ↗🏢LangChain↗🏢Vercel AI↗🏢Mistral AI↗