long-running-server

Long-running agent requests that exceed HTTP timeouts can run in the background using LongRunningAgentServer with Lakebase-backed task persistence. This enables requests to continue executing after the initial HTTP call finishes.

Background polling for async Responses API tasks works via GET /responses/{id} endpoints. Clients request background=true to start the task, then poll the endpoint to retrieve results reliably even when the original request times out.

Yes, streaming long-running agent responses supports cursor-based resumption. By requesting background=true and stream=true, clients can resume streaming using the starting_after cursor and stream event ID transformation ensures consistent response identifiers.

Yes, Lakebase PostgreSQL is required to persist long-running task state. This persistence mechanism ensures results can be retrieved reliably even when the original HTTP request times out or fails.

Handling agent tasks running longer than 120 seconds is best done by replacing AgentServer with LongRunningAgentServer. This approach enables async execution and prevents HTTP gateway timeouts from failing the request.

Agent server requests fail when HTTP clients or gateways limit execution duration, typically around 120 seconds. LongRunningAgentServer solves this by enabling persistent background execution with Lakebase-backed state persistence.