Vercel released AI SDK 7 on June 26, making the TypeScript library’s observability and agent-durability layers production-grade for the first time. The SDK already accounts for over 16 million weekly downloads; the new version adds five feature clusters that target the gap between prototype agents and agents you can actually run in production.
The most consequential change for operators is telemetry. Previous SDK versions required attaching lifecycle callbacks to every generateText or streamText call individually. Version 7 replaces that with a single registerTelemetry() call at application startup that covers all SDK functions automatically. The redesign follows the OpenTelemetry GenAI semantic conventions and integrates out of the box with Datadog, Langfuse, Braintrust, Sentry, and several other observability vendors. Traces now capture the full execution shape of an AI operation: every model call, each tool execution, token counts, errors, and optionally the runtime context passed between steps. For teams billing by usage or debugging tool-execution latency in production, this is the missing layer.
Alongside telemetry, SDK 7 adds WorkflowAgent via a new @ai-sdk/workflow package. The problem it solves is straightforward: when an agentic run spans multiple steps or waits on a human approval gate, a server restart or deployment mid-run currently means starting over. WorkflowAgent checkpoints execution state durably, so runs survive restarts, deployments, and interruptions. It supports streaming, tool approvals, and typed runtime context across workflow step boundaries. This positions the AI SDK more directly against LangGraph and the durable-execution features LangChain and Temporal have been promoting for the past eighteen months.
The version also introduces HarnessAgent, an abstraction for running established coding agents, including Claude Code, Codex, and Pi, through a unified interface. Each harness can be configured with a sandbox environment, custom instructions, and skills; the integration API stays the same regardless of which underlying agent is running. That is relevant for teams building dev tooling where swapping agents should not require rewiring the application layer.
On the developer-experience side, SDK 7 standardizes reasoning configuration across frontier models. Every provider exposes model reasoning differently; the new reasoning option in generateText and streamText maps to provider-native settings from a single field. It also adds per-step performance statistics, including time to first token, output tokens per second, and tool execution latency, directly on the response object. Getting those numbers previously required a custom wrapper.
Two other additions narrow meaningful gaps. A uploadFile API lets agents upload a file once and pass a lightweight provider reference into subsequent model calls, avoiding the overhead of re-uploading the same bytes across a multi-step run. A contextSchema option on tools scopes secrets like API keys to the specific tool that needs them rather than the broader agent context.
The release announcement does not include independent benchmark data on latency improvements or measured overhead from the telemetry layer. Teams evaluating whether to migrate from SDK 6 will need to run their own workload tests; an automated codemod (npx @ai-sdk/codemod v7) handles most of the mechanical migration.
For builders currently managing homegrown observability wrappers around AI SDK, the unified telemetry system is the strongest near-term reason to upgrade. For teams weighing durable agent frameworks, the WorkflowAgent comparison against LangGraph is the evaluation to run before locking in a serverless architecture for late 2026.
Announced on Vercel’s engineering blog on June 26, 2026, by Gregor Martynus.
