We built an agent that fixes production bugs. The hardest part was stopping it.

Description and use

An on‑call agent automatically reproduces failures, writes minimal fixes, and opens PRs, but requires human approval before any write, ensuring safety through strict tool gating and repository isolation.

Problem

On-call engineers spend time manually reproducing bugs, writing tests, patching code, and opening PRs. Manual judgment is required to decide when to write to the repository, leading to either unsafe autonomous bots or slow human review.

Outcome

The agent reduced on‑call engineer effort by automating mechanical steps while maintaining safety through a verified tool‑call firewall, reproduction requirement, and fail‑closed policy. A critical bug in the glob matcher was discovered and fixed by moving the repair service to a separate repository, strengthening the perimeter.

Workflow

• 1. Sentry fires an alert
• 2. Agent reads the issue and culprit frame
• 3. Agent reproduces the failure in a sandbox with a failing test
• 4. Agent writes the smallest patch that makes the test green
• 5. Agent calls a GitHub tool to create a pull request
• 6. TrueForge pauses the turn and presents the exact arguments to the operator
• 7. Operator reviews the diff and approves or rejects the write
• 8. If approved, the pull request is created; if rejected, the agent stops

Evidence

By the time you see that summary, the agent has already decided everything, and you are reading a description of the work written by the same thing that did the work.

Source

If you cannot reproduce the failure, say so plainly, explain what you tried, and stop — do not proceed to a fix.

Source

The agent is required to work six stages in order, and stage three is: reproduce the failure in the sandbox with a test that fails.

Source