Operate AXL
Getting an agent working and keeping it working are different jobs. This section is the second one: what to look at when a run behaves oddly, how to tell whether quality is drifting, and where to start when something has plainly broken.
| If you need to… | Read |
|---|---|
| See what a deployment is doing right now | Use the diagnostics console |
| Follow what happened during a run | Observe runs |
| Check or cancel work that outlives a request | Track asynchronous tasks |
| Know whether the agent is actually any good | Evaluate agent quality |
| Work out why something is failing | Troubleshoot a deployment |
The two questions to instrument early
What a run actually did is answered by observability - traces, tool audit entries, provider calls, and cost, all correlated by run identifier. Set this up before you need it, because the run you most want to inspect is always one that already finished.
Whether it is getting worse is answered by evaluations. A prompt change that improves three examples and quietly breaks twenty is invisible without a dataset, and online evaluation samples real production runs so you find drift without waiting for someone to complain.
Architecture and concurrency invariants live under rather than here.
Next
A live view of what a deployment is doing, and what is wrong with it.
Observe runsTraces, tool audit entries, provider calls, and cost, correlated by run identifier.
Evaluate agent qualityOffline datasets and production sampling.
Track asynchronous tasksTask state and cancellation.
FAQIndexed by the error message you are looking at.