Observability and tracing

Trace every step of a run so that when reliability drops, you can see exactly where and why instead of guessing.

Logs are not enough

A single agent task can involve dozens of model calls and tool invocations. A flat log will not tell you where it went wrong. A trace that captures the full tree of steps, with the inputs and outputs at each node, lets you replay any run and find the exact point of failure.

Instrument to a standard

Recording traces in a common format keeps you portable and future-proof. The OpenTelemetry GenAI conventions define a vendor-neutral way to capture model and agent activity, and platforms like Langfuse give you searchable traces, cost, and latency on top.

Trace in test and in production

The same tracing that helps you debug locally is what lets you understand production behavior. Instrument once and use it everywhere, so a reliability regression is visible rather than mysterious.