Summary: Vercel released AI SDK 7 on June 25, 2026 as an agent runtime with unified reasoning settings, scoped tool context, approval flows, durable execution, timeouts, sandboxing, telemetry, and multimodal integrations. The article recommends adopting approvals, timeouts, and telemetry first because they directly reduce production risk and are measurable after rollout.
With model APIs converged, the hard part of agents has moved to controlling execution: deciding which tool may hold which API key, knowing when a sensitive action must wait for a human, keeping a long-running task alive through restarts, and finding the evidence trail after something breaks. AI SDK 7, released on June 25, 2026, is a systematic answer for that layer.
The Turn from Model Adapter to Agent Runtime
Weekly downloads for AI SDK top 16 million, and the package serves as the standard TypeScript layer spanning model providers; Vercel’s open-source agent framework, eve, is built on that same layer. The stacking choice is itself a signal: once a framework opts to sit on the SDK rather than tie itself to one provider, the ecosystem’s center of gravity has already shifted. Version 7 puts its updates on five fronts — building agents, running agents, plugging in any agent harness, observing agents, and reaching past text into voice and video. Taken together, the five say one thing: this is no longer a wrapper around model APIs but a shared interface for an agent runtime. For product teams, that means the shared execution concerns — permissions, recovery, observability — are starting to get standard answers, and agent code is no longer bound to one vendor’s loop.
The Control Plane: Reasoning Settings, Tool Context, and Approval Flows
Reasoning is the first control to be unified. generateText and streamText accept a single reasoning option that maps to provider-native settings in one line (reasoning: 'high', say), ending the per-provider parameter dialects.
Tool Context is about how finely permissions can be sliced. A growing share of tools comes from third parties independent of the agents that call them, and such tools need inputs an LLM will never produce — API keys are the classic case. Each tool can now declare its own typed context through a schema, and that context reaches that tool alone: a third-party tool can no longer page through the agent’s whole context. For teams shipping tools as products, this boundary is what matters. A separate typed runtime context is readable and modifiable during prepareStep and tool approval functions, letting a run tune prompts or model selection between steps; being handy is not the same as being authorized, and secrets still belong under least privilege.
Approvals take three forms: user-approval on particular tools, per-tool approval functions that either auto-approve or pass the decision to a human, and a catch-all. All three are defined on ToolLoopAgent, generateText, and streamText, so approvals ride the usual execution APIs and require no separate calling convention. Higher-risk workflows can opt into HMAC-signed approvals, which re-verify tool inputs and policies before execution and leave an approval impossible to forge.
The Execution Plane: Durability, Four-Level Timeouts, and Sandboxes
WorkflowAgent, from @ai-sdk/workflow, turns execution durable and resumable: a run continues across process restarts, deploys, interruptions, and approvals that arrive late. The line to draw: what durability restores is process state, not side effects. Emails, payments, publishes — these still call for their own idempotency design, or a recovery could execute one twice.
Agents run out of road in many ways: a provider opens a stream only to stop sending chunks, a tool never returns, a multi-step run eats past its total budget. One timeout can never cover all of that, so version 7 lifts timeouts into a first-class option at four levels — total, per-step, per-chunk, per-tool — with an expiry aborting the run as TimeoutError. The SandboxSession abstraction cuts tool development loose from sandbox choice, so the same tools pair with any sandbox provider, Vercel Sandbox being one example. The benefit is concrete: prove a tool set out in a local sandbox, then carry it to a cloud sandbox for production while the tools themselves never notice the move.
The Integration Plane: One-Shot Uploads, MCP Apps, Terminal TUI, and Harnesses
At the top level, uploadFile and uploadSkill upload files and skills a single time, after which they are reused through a provider reference instead of being re-sent each turn. MCP Apps let MCP servers sort tools into model-visible and app-only groups and render app UIs inside sandboxed iframes — the model’s view and the human’s view can at last travel separately, where these human-facing surfaces previously had to squeeze through chat messages.
In the terminal, @ai-sdk/tui brings an interactive agent up in a few lines through runAgentTUI, and reasoning, tools, and markdown are all supported. The bigger move is the experimental HarnessAgent: one API that drives agent harnesses such as Claude Code, Codex, and Pi, taking a sandbox, instructions, and skills as configuration, with sessions able to pause and resume. The experimental label is explicit, which means the interface still has room to change — budget for that adjustment before adopting.
Observability and Multimodal: The Senses a Production Agent Needs
Telemetry registers once and takes effect everywhere: OpenTelemetry (GenAI conventions) plus ready-made integrations for Datadog, Langfuse, Sentry, and more. On Node.js there is also the ai:telemetry diagnostics channel, emitting structured events for model calls, tools, embeddings, and reranking — one subscription turns the activity into traces, and observability stops being a pile of callbacks each application wires by hand. For an agent running several models and tools at once, hunting bottlenecks finally happens on a single timeline.
Both multimodal pieces carry an explicit experimental tag. Realtime spans browser WebSocket sessions with ephemeral tokens created on the server side, providers OpenAI, Google, and xAI, and a React hook that returns UIMessage[]. generateVideo already has implementations from fal, Google AI Studio, Google Vertex, and Replicate. One unified interface makes swapping easy, but media models differ sharply in capability and pricing, so confirm each provider’s limits before production.
A Sensible Adoption Sequence for Builders
The migration itself is tool-assisted: npx @ai-sdk/codemod v7 together with a migration skill keeps the mechanical cost of the move from v6 low. You are not required to switch everything on at once. A stable order: make approvals, timeouts, and telemetry solid first — they cut production risk directly and are the easiest to check afterward; what approvals intercepted, how many stalls a timeout rescued, which tool telemetry ranks slowest — every one of those is measurable. Bring in WorkflowAgent durability for long-running tasks next; leave harness integration and the voice and video experiments until they stabilize. AI SDK 7 is not valuable for the length of its feature list; it is valuable for moving agents from assembled scripts toward a runtime with permission boundaries, recovery, and observability — the very foundation agents were missing on the way to production.