Vercel AI SDK 7 turns a model-call helper into an agent platform

Vercel shipped AI SDK 7 on 2026-06-25, and the release reads less like a library update than a category repositioning. The AI SDK started as a TypeScript helper for model calls and streaming UI. Version 7 turns it into Vercel’s agent platform layer, with primitives for reasoning, approvals, durable execution, sandboxed tools, harness adapters, MCP apps, and observability. The breaking requirements matter too: Node.js 22 and ESM are now the floor.

What shipped

The core API now speaks in agent concepts. runtimeContext carries orchestration state across steps, tools can declare typed context with contextSchema, and provider file uploads let large inputs be uploaded once and reused by reference. The release also adds provider skill uploads, so Anthropic-style skills can move through provider-managed environments instead of being re-sent as raw prompt material.

Production controls are the bigger shift. Tool approvals can be declared per agent or per call, including policies that require user approval, auto-approve low-risk tools, or deny sensitive operations. WorkflowAgent adds durable execution through @ai-sdk/workflow, giving long-running agents somewhere to persist state across deploys, process restarts, and delayed approvals. Timeouts now cover total run time, step time, chunk latency, and tool execution, which gives teams a way to make agent failure modes operational instead of mysterious.

Vercel also added a harness layer for existing coding agents. Claude Code, Codex, Deep Agents, OpenCode, and Pi can be wrapped behind the same agent interface used by AI SDK-native agents, with sandbox support and Vercel OIDC for AI Gateway. Observability moved in the same direction: @ai-sdk/otel, Node.js tracing channels, lifecycle callbacks, and step performance statistics are all first-class in the release.

Where this lands in the market

The release makes Vercel’s agent stack look less like a collection of separate products and more like a control plane. AI Gateway supplies model routing and spend controls. Sandbox supplies execution. Workflows supplies durability. The AI SDK becomes the developer API that binds them together. That matters because the agent-runtime market is currently fragmented between framework-first players like LangChain, product-first agents like Claude Code and Codex, and platform clouds trying to own deployment.

AI SDK 7’s sharpest competitive move is the harness layer. Vercel is not asking teams to rewrite their agents into a Vercel-only runtime. It is offering a way to wrap the agent tools developers already use, then run them inside Vercel’s sandbox, gateway, workflow, and telemetry surfaces. That is a more credible adoption path than trying to replace Claude Code or Codex outright.

The tradeoff is migration cost. ESM-only packages and Node 22 as a hard requirement are reasonable for new agent apps, but painful for older Next.js monorepos and internal tools still pinned to CommonJS or Node 20. The codemods lower the mechanical cost, but teams will still need to audit semantic changes around telemetry, tool context, streaming helpers, callbacks, and prompt fields.

What’s worth watching

1. Harness adoption. If teams wrap Claude Code and Codex through AI SDK instead of writing bespoke orchestrators, Vercel becomes the agent operations layer even when it is not the agent model provider.
2. WorkflowAgent production use. Durable execution is where demos turn into systems. The next signal is whether Vercel publishes reference architectures for approval-heavy agents, long-running repo work, and human-in-the-loop operations.
3. Node 22 migration friction. The SDK is pushing the ecosystem toward newer runtime assumptions. If app teams stall on Node 20, AI SDK 7 will split adoption between greenfield agent products and legacy app integrations.

The plain read is that Vercel wants the AI SDK to be the React-shaped API for agents: not the model, not the IDE, and not the cloud by itself, but the developer surface that lets all of those pieces snap together. Version 7 is the first release where that ambition is explicit.

Sources cited

1. Vercel: AI SDK 7 is now available vercel.com
2. Vercel AI SDK GitHub release: ai@7.0.0 github.com
3. AI SDK migration guide for v7 ai-sdk.dev