Skip to main content
min read

AI agent observability starts at the tool call, not the chat transcript_

OpenTelemetry’s GenAI work highlights the chain hidden behind one answer: model calls, tokens, retries, and tool invocations. Without trace context, teams blame the model for failures created by an API or retry loop.

  • OpenTelemetry
  • AI Agents
  • Observability

OpenTelemetry’s GenAI work highlights the chain hidden behind one answer: model calls, tokens, retries, and tool invocations. Without trace context, teams blame the model for failures created by an API or retry loop.

A transcript shows what the user saw. It does not show which retrieval query ran, how long a tool waited, what was retried, or where cost accumulated.

What changes in practice_

Instrument model and tool spans, attach stable request and workflow IDs, and redact content by policy rather than recording everything by default.

  • Trace every model and tool boundary.
  • Record latency, tokens, retries, and outcome.
  • Separate operational telemetry from sensitive content.

Our take_

The durable advantage is not adopting the newest tool first. It is building the identity, state, evidence, and operating boundaries that let a real team own the system after launch.

Source_

Our take - not a reprint. Read the original for full reporting.

Want this applied to your stack?

Map your systems or book discovery - we keep humans accountable for what ships.