Enable Signal

In Instrument your agent, you set up tracing and debugged a single run by reading its trace. That works for a handful of test queries. In production, your agent handles far more requests than you can review by hand, and the failures that matter are the ones that keep happening. Signal finds those for you. Enable it on your project and it reviews new traces on a schedule, grouping recurring failures into ranked issues. Each issue is an investigation: an overview of the pattern, links to the traces where it happened, and a proposed fix. Instead of hunting for problems, you start each day with a ranked list of them. This guide takes a few minutes and happens entirely in the Arize AX UI. The next guides take one of the issues Signal finds and turn it into an automated check, then prove a fix works.

​

Turn on Signal for your project

That’s it. Signal runs on its own from here: each review reads the traces that arrived since the last one and files anything recurring as an issue.

​

Read your first issue

Open an issue from the list. Each one is an investigation to review, not a change to merge:

• Overview: the failure pattern Signal found and why it matters.
• Evidence: the traces where it happened. Open them in the trace view to see each failure in context, the same way you debugged a single run in Part 1.
• Proposed fix: a suggested prompt, code, or configuration change to evaluate.

If the list is empty, wait for the first scheduled review and confirm the project has traces. You can continue to the next guide in the meantime and come back once Signal files an issue.

​

Configure Signal

On the Signal tab, open ⋯ → Configure Signal to change how often Signal runs and give it extra context about your agent (what it does, what “wrong” looks like), then click Save. On Enterprise plans, the same dialog can also attach a GitHub repository so Signal opens a pull request for a proposed fix instead of only describing it. Review any pull request before you merge it: Signal never deploys your agent. For plan details, monthly issue limits, and repo setup, see Pricing and usage and the Signal documentation.

​

Congratulations!

Your project now watches itself. Signal reviews every batch of new traces and files recurring failures as ranked issues, so you find out about a pattern from the issue list instead of from a user report. An issue tells you a failure keeps happening. It doesn’t yet score every new response for that failure as it arrives, so pick the issue that matters most (a hallucinated answer, a wrong tool call, an agent stuck in a loop) and carry it into the next guide. Next up: turn that failure into an evaluation, an automated check that scores every new trace, so you can measure the problem continuously and later prove your fix worked.