AI Agents

AI SDK 7 Bets That the Hard Part of Agents Isn't the Model — It's Staying Alive

Published Jun 30, 2026AI Agents

Vercel's AI SDK 7, released June 25, 2026, reorients the popular TypeScript library around production agents — adding durable execution, signed tool approvals, first-class timeouts, and redesigned telemetry to close the gap between a working demo and a system that survives a crash.

What shipped

On June 25, 2026, Vercel released AI SDK 7, a major version of its provider-agnostic TypeScript SDK for building AI applications [1][8]. Vercel says the library pulls more than 16 million weekly downloads and is the same layer its open-source agent framework, eve, is built on [2]. The headline of this release is a shift in purpose: from "stream tokens into a chat UI" toward running production-grade agents, organized around three verbs — Develop, Run, and Observe [2].

The marquee feature is durable execution. A new package, @ai-sdk/workflow, introduces WorkflowAgent, a long-running, resumable agent [4]. The problem it targets is familiar to anyone who has shipped an agent: the standard in-memory loop (ToolLoopAgent) loses all progress the moment the process crashes, redeploys, or restarts mid-run. WorkflowAgent instead runs inside a Vercel Workflow, where each tool execution becomes a durable step with automatic retries, and execution state persists to durable storage between steps [4]. The practical payoff, as independent coverage notes, is durable streams: getWritable() yields a persistent stream that multiple clients can disconnect from and reconnect to — "no Redis or custom pub/sub required" — and an approval request can sit suspended for hours until a human responds, after which the agent simply resumes [7].

The trade-offs are explicit in the docs. ToolLoopAgent lives in the ai package, runs in memory, and offers generate() and stream(); WorkflowAgent lives in @ai-sdk/workflow, survives restarts, retries tools automatically, treats stream() as the primary API, and drops generate() [4].

The rest of the production toolkit

Durability is the lead, but the release rounds out the operational surface. Tool approvals are hardened: tools still run automatically by default, but you can now require approval before execution, and high-risk workflows can opt into HMAC-signed approvals that cryptographically bind the original tool inputs to the approval token — preventing argument tampering between request and resumption [2]. v7 also revalidates tool inputs and policies on replay, and exposes policy-based approvals rather than per-call flags [2].

Timeouts become first-class, with total, per-step, per-chunk, default-tool, and per-tool budgets, and abort reasons that flow through the stream and UI protocols [2]. Telemetry graduated out of experimental — experimental_telemetry is now simply telemetry — with OpenTelemetry moved into a dedicated @ai-sdk/otel package, a registerTelemetry() API, a documented Langfuse integration, Node.js diagnostics-channel tracing, per-step performance stats, and controls that keep secrets out of telemetry by default [2]. Beyond text, v7 adds stable speech and transcription, reranking, multimodal embeddings, and experimental realtime voice and video across OpenAI, Google, and xAI [2]. Finally, harness integrations let teams swap between Codex, Claude Code, Deep Agents, OpenCode, or Pi without rewriting agent logic [2].

Why it matters

Vercel's framing is blunt: everyone can call a model, so the differentiator is keeping the agent loop alive after the demo [2][6]. v7 aims squarely at the operational failure modes — crashes, restarts, runaway loops, leaked secrets, unobservable behavior.

The most consequential change is where durability lives. Until now, TypeScript teams needing crash-surviving agents typically reached outside the SDK. LangChain's own materials observe that the OpenAI Agents SDK does not absorb workflow-level durability natively, leading teams to pair it with Temporal or DBOS [14]. Temporal itself published a guide for adding durability to the AI SDK via a plugin [10], and Vercel's April 2026 durable-execution post previewed WorkflowAgent as a fully native implementation [5]. With v7, that durability — plus resumable streams — moves into the SDK itself, reducing the need to stitch together Temporal, DBOS, or Redis for the persistence use case [7].

Security and governance also move into the framework. Signed approvals, replay revalidation, and default secret redaction land into a tense climate: a DEV Community roundup the same week paired new agent primitives with a run of agent security findings [9], and a Five Eyes warning days earlier flagged near-term AI-enabled cyberattacks [15].

What it changes for practitioners

This is a major version with real breaking changes — renamed telemetry options, Google provider renames such as createGoogleGenerativeAI → createGoogle, and a move to ESM and Node.js 22 — but it ships codemods (npx @ai-sdk/codemod v7) and an agentic migration skill [3]. Reasonable guidance: migrate promptly if you depend on tool context, agent loops, approvals, or telemetry; defer if you only stream text into a stable chat UI [3][6].

Two caveats deserve flagging. First, release status is muddled: Vercel's channels say "now available" and recommend ai@latest [1][2][8], while Alpha Signal describes it as "now available in beta" [7]. Second, portability: WorkflowAgent durability is built on Vercel's Workflow DevKit, whereas the Temporal/DBOS route is more vendor-neutral [4][10]. Sources describe the native integration positively but do not deeply probe lock-in — an open question for teams to weigh.

The competitive lane

Comparison sources place AI SDK alongside LangGraph, CrewAI, PydanticAI, Mastra, the OpenAI Agents SDK, and others [11][12][13][14]. LangGraph offers graph-level control plus the broader LangSmith lifecycle ecosystem; Mastra is a durable TypeScript backend often paired with CopilotKit; the OpenAI Agents SDK favors zero-infra hosted tools but lacks native durability [11][13][14]. v7's distinct contribution is folding native durable execution, in-framework approvals, and telemetry into a single TypeScript lane — not something to be assembled from separate tools.