Observe runs

When a run does something unexpected, you want to reconstruct it: what the agent was asked, what it decided, which tools it called with which arguments, what came back, and what it cost. AXL records all of that, and the job here is mostly making sure it lands somewhere you can read.

Set LOG_LEVEL or RUST_LOG for server logging. Production mode emits structured JSON logs. Keep full wire logging off in production because model requests and responses may contain user content.

Every tool invocation is written to the audit log. Run observers can also export spans and model usage to the configured observability backend, including Langfuse when enabled.

When investigating a run, start with its session or run identifier, then correlate agent events, tool audit entries, provider calls, and any workflow or task state. See Operations reference.

Next

• Track asynchronous tasks - following work that outlives the request.
• Evaluate agent quality - measuring quality.
• Troubleshoot a deployment - when a trace is not enough.