Traditional applications can often be debugged from requests, logs and database state. Agents add uncertainty: why a tool was selected, what retrieval returned, whether a retry changed context and what humans edited. Logging only the final answer makes real failures almost impossible to replay.
1. The short answer
Give every task a trace ID linking the request, prompt version, model version, retrieval, tools, permission results, retries and final state. Record duration and error type for each span, but redact sensitive content instead of writing customer data into logs for convenience.
2. The context
Observability does not mean storing every token. Record events that explain behavior: which tool ran, whether its schema passed, what state it returned and why the agent continued or stopped. For long context, store summaries and evidence references with protected access to raw content.
3. The real trade-offs
Track both success and risk: completion, tool success, P95 latency, duplicate calls, refusal, human takeover, rollback and cost per successful task. Model response time alone hides retrieval, provider and approval delays.
4. How to put it into practice
Error handling should tell operators what to do next. A timeout should query state before retrying; a permission denial should show the boundary; an invalid output should record repair attempts. Alert by user impact and recoverability so low-risk retries do not create equal noise.
5. Common failure modes
Logging only prompts and completions misses tool inputs, outputs and versions. Separating telemetry from business results is also a mistake: a team may know how many calls happened without knowing whether the customer completed the task.
6. FDE perspective
Agent observability should be designed around delivery of the task, not around the words the model produced. A complete trace shows whether the fault came from data, model, tool or permission and gives automation evidence for improvement.
A practical checklist before shipping or buying
- Give each task a replayable trace ID.
- Record tools, permissions, retries and versions, not only final text.
- Measure delivery, latency, risk and cost together.
- Redact logs and alert by impact and recoverability.
What to do next
Add traces, tool events and business completion to one agent workflow before expanding monitoring.
This article is based on public documentation and research from OpenTelemetry documentation. The FDE perspective is editorial analysis, not a quotation from the source.
Turn this read into a practical next step.
Matched to this article automatically; each tool is managed independently from the admin.