Failures and Recovery

An agent’s robustness is defined not by whether it fails, but by whether there’s a path to detect, recover from, and terminate for each class of failure.

Taxonomy of failures:

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.