Ephemeris.
Today the agents got somewhere to run — a runtime, a key it borrows instead of keeps, a lab bench, and a mirror held up by Anthropic.
The agent harness is open source now.
Vercel released eve, an agent framework that ships the production scaffolding you keep rebuilding by hand. You write what the agent should do; eve runs it like infrastructure — and the whole thing is open.
Agents cold-start in ten milliseconds.
Cloudflare cracked the Agents SDK open into a runtime any framework can build on — durable execution, sandboxed code, a virtual filesystem. Flue, from the Astro team, is the first to take it up; you deploy agents from the dashboard in a few clicks.
Your agent borrows the keys instead of keeping them.
Vercel Connect retires the long-lived provider secret. Apps and agents prove identity over OIDC and get a short-lived, task-scoped token at runtime — access requested per action, never stored. Eight integrations ship today.
| Provider | Credential | At rest |
|---|---|---|
| Slack · GitHub | OIDC → scoped token | none |
| Linear · Discord | OIDC → scoped token | none |
| Notion · Salesforce | OIDC → scoped token | none |
| Figma · Snowflake | OIDC → scoped token | none |
| Resend · MS Teams | in development | — |
A model ran the reaction until it worked.
OpenAI and Molecule.one put a near-autonomous AI chemist, built on GPT-5.4, to work on a stubborn step in a drug synthesis — and it improved the reaction. The model proposed conditions, read the results, and tried again, closing a loop a bench chemist normally runs by hand.
It proposed the conditions, read the results, and tried again — the loop a bench chemist runs by hand.
Grading models on the lab work scientists actually do.
OpenAI released LifeSciBench, an evaluation for how AI systems handle real-world life-science research tasks rather than textbook recall.
The benchmark targets the messy, multi-step problems a working scientist faces — protocol design, interpretation, the judgment calls between steps — not multiple-choice trivia a model can pattern-match.
If you build evals, the method is the lesson: domain-grounded grading where subject-matter experts both write the questions and check the answers, so a high score means competence rather than recall.
Only 15% of Americans trust the labs.
Anthropic published the first Anthropic Public Record, a survey of how the broader public — not just AI users — feels about AI. It's a baseline worth knowing if you ship AI products, and the trust gap is the headline number.
That's today.
Six picks from a day when agent infrastructure went to production — a runtime, a credential broker, a framework — alongside OpenAI pushing models into the lab and Anthropic opening its books on public trust.