Failures and Recovery

An agent’s reliability is defined not by whether it makes mistakes, but by whether there’s a path to detect, recover from, and terminate for each class of error. Taxonomy: API level (rate limits, timeouts), tools (hallucinatory calls, repeating the same error), context (overflow, structure corruption), control flow (dead loop, death spiral). Recovery with increasing transparency: silent retry with exponential backoff → degradation and continuation → show user only after exhausting automatic means. Each path has a fuse with a threshold from real data.

Related: [Five Functions of Harness], [Event-Driven Asynchronous Agent]