About this session
Amara shares the incident log from Openfield's first quarter running support agents with real customers. Runaway tool loops, stale context and one memorable refund bug each get a post-mortem and a fix. The talk is a field guide to guardrails, budgets and the unglamorous monitoring that keeps agents honest.
What you'll take away
- The five failure modes that caused most of our incidents
- Budgets, timeouts and kill switches that actually trigger
- What to log so a bad conversation is debuggable at 2 AM