CH 13 · Reliability: Retries, Idempotency, and Failure Handling · 08/13
Read as one pageFailing within a stream
Because agent responses are streamed (Chapter 8), failures often occur after you've started responding. Handle it:
- Wrap the orchestration so that on error you can still emit an
errorevent and a terminating sentinel, so the client shows a clean failure rather than hanging. - Don't lose persisted progress: the pre-saved user message and any incrementally-saved results survive.
- Clean up server-side resources (abort the in-flight model call) when the client disconnects, so abandoned turns stop burning tokens.
Finished this idea?
Your learning progress stays in this browser.