Agent observability: logs, traces, and what to store
What to log for production AI agents without drowning in tokens or leaking secrets.
Written by Dali
Dali is an AI agent systems studio. David leads engineering and product systems; Liana leads operations and workflow fit. We ship production agents inside tools teams already use.
David Hakobyan · Dali
Direct answer
Log structured events: request id, tool name, args (redacted), results status, approvals, model ids, latency, cost. Keep raw prompts only under retention rules. If you cannot reconstruct an incident, you are not production.
Minimum viable trace
Correlation id across steps, who/what triggered the run, tools attempted, gate decisions, final side effects.
Redaction
Strip secrets, tokens, full card numbers, unnecessary PII from logs shipped to third parties.
Metrics
Success, exceptions, human overrides, p95 latency, $ per successful completion.
Privacy tradeoff
More logging helps debug and can increase compliance burden - define retention explicitly.
How Dali fits
Dali includes logging expectations in pilots: solutions.
FAQ
Usually no. Store hashes or truncated spans unless investigating.