Vercel releases AI SDK 7 with approvals, durability, and telemetry

• Vercel positioned AI SDK 7 as a production-focused release, with Vercel's launch post naming approvals, durability, and telemetry as the foundation pieces for agents and AI platforms.
• Early adapter feedback from threepointone's migration note said the release brings breaking changes, enough that work on Agents SDK support is being delayed until there is a cleaner migration path.
• Around the release, Vercel Dev's coding-agents post and ctatedev's eval screenshot highlighted a design pattern Vercel is pushing hard: skills, linters, evals, and runtime assertions living in the same agent stack.
• Observability also got first-class surface area, with Vercel Dev's Workflows trace viewer post showing search across spans plus step-level inputs, outputs, and metadata.

You can jump from Vercel Dev's coding-agents writeup to the redesigned trace viewer, and community posts like omarsar0's eve walkthrough and ctatedev's eval example make the release feel less like a model SDK bump and more like Vercel standardizing an agent runtime around tests, traces, and approvals.

Vercel's own framing was blunt. The launch post called AI SDK 7 the release that sets the foundation for production agents, and the named primitives were approvals, durability, telemetry, and more.

The surrounding commentary points in the same direction. ctatedev's summary described the SDK as an AI layer spanning Next.js, React, Vue, Svelte, Expo, TanStack Start, Node.js, and more, now with deeper agent infrastructure.

The first useful caveat came from downstream maintainers, not the launch copy. threepointone said AI SDK 7 looks "very very good," but also said the breaking changes were significant enough that support would not land in Agents SDK immediately.

In the same post, threepointone's note said the team wants to bundle the migration with other dependency changes, possibly including a new MCP SDK. A follow-up reply from the same thread said earlier migration attempts had been "a bit hairy."

The clearest implementation detail around the release is Vercel's preference for keeping agent scaffolding close to the runtime. Vercel Dev said its internal coding-agent system uses a skill, linters, evals, and an updating loop.

ctatedev's eval screenshot shows the style directly. The eval script asserts tool usage, checks that a dangerous tool was not called, and uses a judge step for reply quality, all in TypeScript against the real agent runtime. omarsar0's eve walkthrough made the same point from the outside, saying everything is files, including tools, skills, and evals.

Observability shipped with visible UI, not just API language. Vercel Dev's trace viewer post showed span search, zoom, and step inspection with full inputs, outputs, and metadata for Workflows.

The ecosystem motion started immediately. Linq's adapter announcement thread said Linq became a vendor-official adapter for Vercel's Chat SDK, spanning iMessage, RCS, and SMS, while rauchg's Cursor post hinted at the SDK showing up in coding surfaces beyond Vercel's own stack.