AI SDK 7.0.154 lets running agents accept steering messages and close failed steps
Released October 26, AI SDK 7.0.154 adds experimental steering so an application can queue a user message while an agent is running and deliver it at the next step boundary, plus lifecycle callbacks for no-output streams and later provider failures.
Why it matters
Generative interfaces need to represent corrections, cancellation and changed intent during long tasks while reliably closing loading states, step traces and telemetry on failure. This release moves mid-run input and failed-step closure into the core loop instead of forcing each frontend to invent its own queue.
A closer look
experimental_createSteering works with generateText, streamText, ToolLoopAgent and createAgentUIStream. steer(message,{id}) inserts a separate user message at the next boundary after the current step and its tools finish; a waiting message can continue a loop that would otherwise end, and retrying the same id returns the first outcome. Outcomes distinguish delivered from finished, stop-condition, waiting-for-tool-results, aborted and error. onStepError closes a started step that produced no result and reports callId, provider, modelId, stepNumber and error; telemetry events also carry aiVersion.
Who this is for
Frontend and platform teams building steerable generative UI, long-running agents, approvals, persisted chat history and step telemetry
Original sources
- Vercel AI SDK official release and commits ↗github.com
- Experimental steering implementation and tests ↗github.com
- Failed-step lifecycle callbacks and telemetry closure ↗github.com
- AI SDK version in telemetry events ↗github.com
Revisions & corrections (1)
— Official AI SDK release and signed commits confirm experimental mid-run steering, idempotent delivery outcomes, failed-step callbacks, telemetry versioning, and documented unsupported paths. It is a new release distinct from 7.0.153.