Vercel ships AI SDK 7 for agent builders
AI SDK 7 landed June 25 with durable agents, tool approvals, MCP Apps, and a shared API across Claude Code and Codex. It is also ESM-only, a breaking change.
Vercel released AI SDK 7 on June 25, the framework now pulled in 16 million times a week. The update is aimed squarely at people shipping agents rather than chat boxes, and it adds production plumbing in a few places that have been painful to build by hand.
The headline pieces: a WorkflowAgent whose runs are durable, meaning a long task survives process restarts, deploys, and interruptions; tool approvals that put a human in the loop, with an HMAC-signed option for untrusted callers; and per-step, per-tool, and per-chunk timeouts so a runaway agent does not hang. A SandboxSession abstraction standardizes running agent commands in an isolated environment. There is also support for MCP Apps, which lets an MCP server hide tools from the model and render its own UI in a sandboxed iframe, plus a HarnessAgent that wraps Claude Code, Codex, and similar coding harnesses behind one API.
The catch is a breaking change: every package is now ESM-only, so code that still uses require() has to move to import. Vercel ships a codemod to automate most of the v6 migration.
What it means for you
Durable execution and tool approvals are exactly the controls an agent needs before it touches real systems, so this narrows the gap between a demo and something you would run. If you are weighing agent frameworks, pair this with how enterprise teams are locking down MCP access and the basics in coding agents for non-developers. The sandbox and approval features matter most when you are trying to keep a coding agent from running attacker code.