Vercel shipped AI SDK 5.0 with a first-class agent loop primitive. The new runAgent API takes a tool set, a model, and a termination condition and handles the plan-act-observe loop that every team has been re-implementing on top of streamText, badly, since 2024. Tool definitions stay in the same Zod-schema format the SDK already uses for structured output, which is correct: the easiest migration is the one that doesn’t require you to relearn type definitions you already memorized.
The interesting design choice is that the loop is fully resumable. Each step persists to a pluggable store, and a paused or crashed agent resumes from the last completed step with full context. That moves the SDK from “nice for demos” to “actually plausible in production,” and it puts immediate pressure on the bespoke agent frameworks, LangGraph and the various OSS routing libraries, that were the only resumability option until now. The agent-framework ecosystem just got noticeably smaller.
The trade-off is provider lock-in to the SDK’s tool-call format. Teams already running LangChain or building on raw provider SDKs get less from the upgrade and probably more from staying where they are. For greenfield Next.js apps, the SDK is now closer to a complete agent runtime than any other framework in the ecosystem, which is also probably the point.