Source: github.com/vercel/ai/releases/tag/%40ai-sdk%2Fworkflow%402.0.105
The release adds a toolApproval configuration to WorkflowAgent, set in the constructor or per generate() or stream() call. It takes precedence over a tool's needsApproval setting, and calls it denies receive an execution-denied tool result without pausing the workflow. A new toolOrder option lists tools to send first, with the remaining tools following in alphabetical order.
Timeouts can now be passed as an object with totalMs, stepMs, firstChunkMs and chunkMs, in addition to a number. Invalid values are rejected with an InvalidArgumentError, and a timeout ends a model step even when the provider stream ignores the abort signal. experimental_refineToolInput is also supported in the constructor and per call.
Several fixes change behavior. An abort with a custom reason now calls onAbort instead of onError, while generate() still rejects with that reason. Async generator tools use their last yielded value as the result, and provider warnings are logged through the AI SDK warning logger, which the AI_SDK_LOG_WARNINGS setting controls.
Why it matters if you build agents
Agents built on WorkflowAgent can now gate tool calls with toolApproval without pausing the durable workflow, and set tool ordering and per-step timeouts directly. Code that depends on onError firing for custom-reason aborts should be checked against the new onAbort behavior. Provider warnings will also appear in logs through the AI SDK warning logger.
Get stories like this every weekday
One short email with the day's agent tooling news.