// I was young and needed to ship.

This is a great question! HarnessAgent runs on the host when supported (Pi, OpenAI Agents SDK) and only uses the sandbox for untrusted code. This is the preferred architecture but most SOTA harnesses don't support it yet

This is number 4: a unified abstraction for all agent harnesses. Write the agent once. Run it on any harness without vendor lock-in vercel.com/changelog/program… Huge thanks to @felixarntz for pushing this over the line!

Write your agent once. Run it on any harness. I have high conviction that this will change how we build agents forever. We're sharing an early release today so we can build that future together: - No harness lock in - No model lock in - Full programmability - Strict security

In what might have been the finest bit of serendipitous curation in the history of multicellular life I first read Nick Lane's "The Vital Question, Energy, Evolution, and the Origins of Complex Life" and then Andy Weir's "Project Hail Mary". This can only be explained by intelligent design.

Capitalism FTW

Ohhhh, programmatic control over AI token budgets is actually really useful. Example use case: - start with a more conservative budget - allow re-up with business justification via a slack bot

Huge thanks to the @AnthropicAI team for investigating `just-bash` with Mythos as part of Project Glasswing There are no serious findings. I'll issue a release with minor hardening improvements after my vacation. This is an interesting case for 2 reasons: 1. `just-bash` is a very ripe surface (it's a full implementation of bash and common utilities, bundles QuickJS, CPython, has optional filesystem access) 2. `just-bash` itself was largely written by Opus 4.5 with minimal human review (but deep hardening loops and very paranoid machine-enforced coding rules) justbash.dev

You can just ship things. youcanjustship.vercel.app

Claude Fable 5 is now on AI Gateway. A Mythos-class model for your hardest unsolved problems. 𝚖𝚘𝚍𝚎𝚕: '𝚊𝚗𝚝𝚑𝚛𝚘𝚙𝚒𝚌/𝚌𝚕𝚊𝚞𝚍𝚎-𝚏𝚊𝚋𝚕𝚎-𝟻' vercel.com/changelog/claude-…

I reworked tool approvals in AI SDK 7 so you can combine any tool approval strategy with any tool (approvals are independent of tool definition)

In the next version of Zerolang Agents get closer to the compiler Instead of making agents edit source text, then recover meaning through format/check/build/test loops Zerolang makes the compiler's semantic graph the program Agents query it, patch it, and get checked edits

Vercel AI Gateway recovers on average over 1T tokens a month 🤯 Much like Stripe recovers revenue with smart retries on failed payments or credit card updates. And we do it with 0️⃣ zero markup over the labs; adding redundancy, zero-data retention enforcement, observability, usage APIs, caps, … vercel.com/ai-gateway

A few weeks ago I silently released the skills[.]sh API with "email me" instructions on the docs Since then, my email inbox has received >1200 requests for an API key I didn't want to keep generating API keys via script, but also didn't want to build a brittle "sign in experience" just for API keys Happy to see another beautiful use for short-lived credentials that come out of the box with Vercel!

A drive for every agent