← ELI5 · Nestor G Pestelos Jr

AI Operations · Picture Book

How Can We See What an AI Agent Did?

Give every journey a tracking number.

1. An agent makes many stops

One guest message can visit memory, property data, a model, a tool, and a host.

GuestContextlook-upModelTool orhost

2. Scan the parcel at every stop

A trace records which version ran, what source it used, what action it proposed, and what happened.

TRACKING: journey-4821requestsourcepolicyoutcome

3. Record receipts, not secret thoughts

Keep the evidence and safety decision; do not collect a private diary of the model's hidden reasoning.

KEEPSource IDPolicy ruleTool receiptDO NOT ASSUMEHidden thoughts arean audit log

4. The trace answers “where did it break?”

You can separate a weak model from stale property data, a bad policy, a tool error, or a missed handoff.

ModelProperty dataSTALEFix the broken layer, not everything.

Back to top

Longer version: AI Agent Observability reference entry

Sources: W3C Trace Context; OpenTelemetry GenAI conventions; OWASP Logging Cheat Sheet.