Most teams arrive at agent observability with an APM mindset: latency, error rate, throughput. Those tell you the system is up. They cannot tell you it is right — an agent that confidently returns a wrong answer in 200ms is a green dashboard and a support ticket.
The gap in conventional monitoring
The common failure pattern is tools that trace what happened without evaluating whether it was correct, monitor infrastructure metrics without measuring output quality, and surface failures only after users report them. Agentic systems need the evaluation layer wired into the trace, not bolted on after.
What a useful trace contains
- Structured spans across the whole loop — each LLM call, each tool invocation, each memory read and write.
- Parent-child relationships preserved through handoffs, so a multi-agent run reads as one tree rather than several disconnected traces.
- The actual context window sent, not just the user's message. Most wrong answers are context problems, and you cannot diagnose them without seeing what the model was given.
- Tool arguments and returns, redacted for PII but complete in structure.
- The decision not to act — refusals and no-op turns are signal, and they are usually unlogged.
Scoring production traces
Tracing gives you the raw material; evaluation gives you the verdict. Score a sample of live traces against the same rubrics as your pre-deployment suite, and route the failures back in as new cases. Without that loop, the eval suite ages out of relevance within weeks.
Sample, do not score everything
Scoring every production trace with an LLM judge is expensive and usually unnecessary. Sample continuously, and score exhaustively for the narrow slice where correctness is expensive to get wrong — anything touching money, permissions, or irreversible actions.
Cost and token accounting belong in the trace
Agent cost is emergent: it depends on how many loops a task took, which is a function of prompt quality and tool design rather than anything you set in configuration. Attribute tokens per span so a cost regression points at the step that caused it. Teams routing most traffic to smaller models need this to verify the routing is actually happening.
Alert on quality, not just availability
- Task-completion rate falling for a specific intent.
- Tool-call error rate rising — usually a schema or upstream change.
- Mean loop count per task climbing, which is a cost problem before it is a quality problem.
- Refusal rate collapsing to near zero, which usually means the agent stopped declining things it should decline.
Frequently Asked Questions
Can I use my existing APM for this?
Partly. Distributed tracing gives you the span structure; what it lacks is the evaluation layer that scores whether a span's output was correct. Most teams pair an APM with a dedicated LLM evaluation tool.
How long should traces be retained?
Long enough to investigate a regression after a model change, which in practice means weeks rather than days. Retain the failures indefinitely — they are your eval suite.
What about PII in traces?
Redact at capture, not at query time. Full prompts are the most useful part of a trace and the most likely to contain user data, so the redaction has to happen before it is written.
References
- Agent Observability: The Complete Guide for 2026 — Braintrust
- Top 8 AI Agent Observability Platforms for 2026 — Confident AI
About Jishu Labs
Jishu Labs is a software development company founded in 2016. We build custom software, AI/ML systems, and full-stack web and mobile applications for clients, and we make eight AI tools for software teams.