How to sort, label, and reproduce reported issues with an agent.
Ask the agent to survey open issues¶
You can just ask the agent to look over issues, it will figure out how to use
gh or whatever it needs:
Categorize all open issues. Highlight ones that are easy to close, and bugs that you can reproduce
Then close the closable issues by hand (such as with the gh tool), respond,
etc. The agent can help with any of this if you wish, but keep a human touch on
interactions.
Ask for a report with the categorized issues to guide your next steps.
To scale up, have it launch subagents to fix the bugs it was able to reproduce, opening a draft PR for each.
A specific issue works too:
Is #123 still broken?
MWEs and tests¶
Agents are great at producing a minimal working example (MWE) of some bug you are seeing. They can also adapt an MWE, or other things in an issue, into your test suite as a regression test. See Writing tests for more on this.
Fully automated?¶
It might seem really attractive to have a fully automated system, but it’s a bit early for that. Attempts have been hacked by malicious issue titles with prompt injection (instructions to an agent in the issue title). Comments can hide prompt injection too. Do not run reporter-supplied text through an agent that can act without human review.