> For the complete documentation index, see [llms.txt](/llms.txt).
> A full single-fetch corpus is available at [llms-full.txt](/llms-full.txt).
---
title: Command Center
description: Turn failed deployments, runs, and online evals into issues, each with an evidence-based report, the likely cause, and a fix prompt for your coding agent.
last_verified: 2026-10-29
---

**Command Center** watches a project for failures and turns them into **issues**. For each issue, an agent reads the evidence and writes a report: what failed, the likely cause, the code involved, and a fix prompt you can paste into your coding agent. You fix the problem, then resolve the issue.



## Three kinds of issues

| Kind | Raised when | Example |
| --- | --- | --- |
| **Deployment** | A deployment fails, for example its worker can't start. | A missing package or environment variable. |
| **Runtime** | A run raises an error. | `TimeoutError` in a workflow step. |
| **Behavioral** | A run finishes but fails an [online eval](/docs/improve/online-evals.md). | A scorer judges the output off-topic. |

## Repeats join one issue

A failure with the same cause in the same component joins its existing issue instead of opening a new one, so a failure that hits a hundred runs is one issue with a run count. When a repeat happens on the same deployment as an existing diagnosis, the issue reuses that diagnosis instead of investigating again. A new deployment, a recurrence after you resolved the issue, or a report you marked incorrect starts a fresh investigation.

## What a report contains

The investigator only reads evidence for the failure it's looking at:

- **Deployment issues:** the deployment's status, lifecycle events, and logs.
- **Runtime issues:** the run summary, its events, logs, and trace.
- **Behavioral issues:** the run's input, output, and trace, and each failed scorer's verdict. Scorer results are treated as claims to check against the run, not as the cause.

With **source analysis** on, the investigator also reads the code from the exact commit that's deployed and points to a file and line, marked as either where the error happened or its cause.

The report states the likely cause with its supporting evidence. When the evidence isn't enough to name a cause, the report says what's missing instead of guessing. The fix prompt carries the failure, the decisive evidence, the cited code, and how to check the fix.

## Resolve, reopen, and review

Command Center never closes an issue on its own. Resolve it once your fix ships. If the same failure happens again, the issue reopens.

You can also review a report as correct or incorrect. Marking it incorrect makes the next matching failure start a new investigation, and your review is passed to the investigator as context.

## Turn it on

Enable Command Center for a project from a linked project directory:

```bash
agnt5 command-center enable
agnt5 command-center status
```

It analyzes failures that happen after you turn it on, not earlier ones. Add `--source-analysis` to let it read your deployed code. [Command Center commands](/docs/cli/command-center.md) covers every flag and the settings it reports. Open the project in Studio and select **Command Center** to see issues and reports.

## Next steps

- [Online evals](/docs/improve/online-evals.md): score live runs so behavioral failures reach Command Center.
- [Deploying](/docs/run/deploying.md): deploy, verify, and inspect runs, traces, and logs.
- [How AGNT5 works](/docs/how-agnt5-works.md): runs, checkpoints, and the build, run, improve loop.