Agent observability
OperationsLLM observability · agent tracing
Instrumentation that shows what an agent did and why: prompt, model response, tokens, tool calls, latency, cost and outcome for every step. Without it you cannot audit a decision, and you cannot tell whether the agent got worse this week.
Agent observability extends the three classic signals (traces, metrics, logs) to agentic behaviour: every reasoning step, tool call and agent-to-agent handoff becomes a span you can correlate with the rest of the stack. The de facto standard is OpenTelemetry's GenAI Semantic Conventions, with attributes such as gen_ai.request.model, gen_ai.usage.input_tokens and gen_ai.operation.name; Datadog announced native v1.37 support in December 2025. The term has established public usage across the observability ecosystem. What is Implementa's own editorial position is treating it as a precondition rather than a nice-to-have before you sell autonomy to a client: if you cannot show the trace, you cannot defend the decision.
How it differs from
- Infrastructure monitoring
- Measures CPU, latency and service errors; it does not explain the agent's decision.
- Agent evaluation
- Observability records what happened in production; evaluation judges whether it was right.
Examples
- Reconstructing why a billing agent applied a discount: prompt, retrieved document, tool called, response.
- Alerting when average cost per run jumps 30% after a model swap.
FAQ
- Aren't model logs enough?
- No. Without correlation across steps, tools and the business transaction you cannot rebuild the decision chain.
- Is there an open standard?
- Yes: OpenTelemetry's GenAI Semantic Conventions, currently in beta and already supported by commercial platforms.
Related terms
Sources & further reading
OpenTelemetry — Inside the LLM Call: GenAI Observability with OpenTelemetry · 2026-05-14 · E3