AI SDK Production Playbook

Look forPin ai and every @ai-sdk package to exact versions from one minor line. AI SDK 7 needs Node.js 22 or later and ESM; Vercel still publishes patches on the 5 and 6 lines, so a team can finish a launch before taking a major. Name an owner for upgrades.

Look forAI Gateway is the default global provider, so a plain string such as anthropic/claude-sonnet-5.5 goes through Vercel with gateway billing. Import a provider package with your own keys when traffic must go direct, or set gateway budgets and data controls first; the AI Gateway guide covers both.

Look forSet stopWhen explicitly. ToolLoopAgent defaults to 20 steps, while WorkflowAgent has no default limit. Add AI SDK 7 timeouts per run, step, chunk, and tool. Pass the request's abort signal. One developer calls its absence the most common production bug. A 2026 issue showed generateText returning normally after a mid-loop abort.

Look forSet toolApproval on each tool. Require user approval for data changes and use approval functions for policy checks. Enable signed approvals on high-risk tools and test resume. A June 2026 issue broke every signed approval on resume until a fix shipped. On October 23, 2026 Vercel shipped a first-party web fetch tool, @ai-sdk/tool-web-fetch, exported as experimental_createWebFetchTool and requiring Node.js 22.13 or later. Vercel's docs warn that fetched pages are untrusted and can contain prompt injections. Domain limits are opt-in and requireUrlInUserMessage defaults to false, so set allowedDomains or turn on requireUrlInUserMessage, and require approval when the agent can reach private data. The tool blocks private addresses, runs no JavaScript, and rejects PDFs. Before October 22, 2026, toolChoice: 'none' did not stop a tool call the model returned anyway. Versions ai 7.0.149 and 6.0.314 (October 22) and 5.0.281 (October 23) reject such calls before local execution and withhold streamed tool parts. Teams pinned below those versions should upgrade.

Look forToolLoopAgent keeps state in memory. For runs that outlast a request or wait for approval, use WorkflowAgent on Workflow DevKit. Another option puts SDK calls inside Temporal workflow steps, as Output.ai does. Check the host's function timeout against the longest expected run.

Look forSet experimental_throttle on useChat and test long answers with many tool calls on slower machines. Developers report re-render loops when useChat's messages feed a useEffect, and one team left the SDK for a native implementation after slow UI hooks in a production app.

Look forRegister telemetry once at startup with @ai-sdk/otel, which follows the OpenTelemetry GenAI conventions, and send it to your tracing tool. Decide whether spans carry prompt and response content before traces leave your network.

Look forUse the test model helpers in ai/test for unit tests, and run a small eval per provider before switching. Public GitHub issues show failures specific to one provider. These include Anthropic rejecting valid Zod schemas, reasoning blocks lost across tool steps, and a Bedrock structured-output change that moved one team's failures from 0 to about 10 of 20.