AI Agents
Agent observability — what to log
When the agent does the wrong thing, you have hours (not weeks) to understand why. Log accordingly.
Published by CloudFixies · 30 December 2025
The minimum trace
Every turn: input, tool calls (with arguments), tool responses, model output, latency, tokens, user identity. Everything hashed if it contains PII. Retained for 90 days, then aggregated.
What to skip
Full document contents. Verbose model reasoning that isn’t action-relevant. Anything that recreates the PII you just went to the trouble of not storing.
◆ Takeaway
Log tool calls and outcomes, not stream-of-consciousness reasoning.