Once model APIs are unified, the hard part of agents moves to execution control: which tool gets which API key, when sensitive operations need human approval, how long-running tasks survive restarts, and where to look when something breaks. AI SDK 7, released on June 25, 2026, is a systematic answer to that layer.
From Model Adapter to Agent Runtime
AI SDK passes 16 million weekly downloads as the TypeScript SDK layer that works across model providers, and Vercel’s open-source agent framework, eve, is built on top of it — itself a signal that when a framework chooses to sit on the SDK layer rather than bind to a single provider, the ecosystem’s center of gravity has already moved. Version 7 concentrates its update on five fronts: developing agents, running agents, integrating any agent harness, observing agents, and going beyond text into voice and video. Read together, the message is clear: the SDK is no longer positioning itself as a wrapper around model APIs but as a shared interface for an agent runtime. For product teams, that means the common execution concerns — permissions, recovery, observability — are starting to have standard answers, and agent code no longer has to be welded to one vendor’s loop.
The Control Plane: Reasoning, Tool Context, and Approvals
Reasoning control gets unified first. A single reasoning option on generateText and streamText maps to provider-native settings in one line (for example, reasoning: 'high'), so you stop maintaining a parameter dialect per provider.
Tool Context addresses permission granularity. Tools are increasingly developed by third parties independently of the agents that use them, and they need inputs an LLM never generates, such as API keys. Each tool can now declare its own typed context via a schema, and only that tool receives it — a third-party tool can no longer read the whole agent’s context, which is the boundary that matters when tools are shipped as products. A separate typed runtime context is readable and modifiable during prepareStep and tool approval functions, letting later steps adjust prompts or model selection; convenience is not authorization, and secrets should still follow least privilege.
Approvals come in three shapes: user-approval for particular tools, per-tool approval functions that can auto-approve or forward to a human, and a catch-all. All three are defined on ToolLoopAgent, generateText, and streamText, so approval rides the same execution APIs instead of demanding a separate calling convention. For higher-risk workflows, opt-in HMAC-signed approvals re-verify tool inputs and policies before execution, so an approval cannot be forged.
The Execution Plane: Durable Workflows, Timeouts, and Sandboxes
@ai-sdk/workflow’s WorkflowAgent makes execution durable and resumable: runs survive process restarts, deploys, interruptions, and delayed approvals. One line worth drawing clearly — durability resumes process state, not side effects. Emails, payments, and publishes still need their own idempotency design, or a recovery may execute them twice.
Agents stall in different ways: a provider opens a stream and stops sending chunks, a tool hangs, a multi-step run exceeds its total budget. A single timeout knob never fits, so version 7 promotes timeouts to a first-class option across four levels — total, per-step, per-chunk, and per-tool — with aborts surfacing as TimeoutError. A SandboxSession abstraction decouples tool development from sandbox choice, letting the same tools run with any sandbox provider, such as Vercel Sandbox — the payoff being that a tool set can be validated in a local sandbox and then moved to a cloud sandbox for production without the tools knowing the difference.
The Integration Plane: Uploads, MCP Apps, TUI, and Harnesses
Top-level uploadFile and uploadSkill APIs upload files and skills once and reuse them by provider reference instead of re-sending them every turn. MCP Apps let MCP servers split tools into model-visible and app-only categories and render app UIs inside sandboxed iframes — what the model sees and what a human sees can finally be separated, where these human-facing surfaces previously had to be smuggled through chat messages.
For terminals, @ai-sdk/tui runs an interactive agent in a few lines via runAgentTUI, with support for reasoning, tools, and markdown. The bigger step is the experimental HarnessAgent: one API to run agent harnesses such as Claude Code, Codex, and Pi, configurable with a sandbox, instructions, and skills, with sessions that can pause and resume. Because it is explicitly experimental, expect the surface to keep moving and budget for adjustment before adopting.
Observability and Multimodal: Eyes and Ears for Production
Telemetry registers once and applies globally: OpenTelemetry (GenAI conventions) plus ready-made integrations including Datadog, Langfuse, and Sentry. On Node.js, an ai:telemetry diagnostics channel emits structured events for model calls, tools, embeddings, and reranking, so a provider can subscribe once and convert activity into traces — observability stops being a pile of hand-wired callbacks. For agents running many models and tools at once, bottleneck hunting finally gets a single timeline.
Both multimodal additions are explicitly experimental. Realtime support covers browser WebSocket sessions with server-created ephemeral tokens, providers including OpenAI, Google, and xAI, and a React hook that returns UIMessage[]. generateVideo ships provider implementations for fal, Google AI Studio, Google Vertex, and Replicate. The unified interface makes switching easy, but media models vary widely in capability and pricing — verify each provider’s limits before production.
An Adoption Order for Builders
Migration has tooling: npx @ai-sdk/codemod v7 plus a migration skill lowers the mechanical cost of moving from v6. You do not need to adopt everything at once. A sensible order: land approvals, timeouts, and telemetry first, because they directly reduce production risk and are the easiest to verify afterward — what approvals intercepted, how many hangs a timeout rescued, and which tool telemetry shows as slowest are all measurable improvements. Add WorkflowAgent durability for long-running tasks; evaluate harness integration and voice or video once they stabilize. The value of AI SDK 7 is not the length of the feature list but the shift from ad-hoc agent scripts toward a runtime with permission boundaries, recovery, and observability — precisely the foundation agents need to reach production.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.