Engineering
Bug reports that arrive reproduced
An engineering team of eight lets a Bug Reproduction Agent turn vague reports into Linear issues with steps, logs and a suspected file.
· 5 min read
"It doesn't work on my phone" is not a bug report. An eight-person engineering team put a Bug Reproduction Agent between the report channel and the backlog.
From Slack to a real issue
Reports land in a Slack channel. The Agent reads each one, asks the reporter the missing questions in the thread, and when it has enough writes a Linear issue with steps, expected and actual behaviour and the environment.
Reading the code
With GitHub connected, it searches the repository for the code path involved, links the likely file and recent pull requests that touched it, and adds a short hypothesis. It does not open pull requests; engineers do.
Keeping the backlog honest
Duplicates are linked instead of created. Issues without a reply from the reporter for three days get a nudge. The on-call engineer gets a daily summary of what came in.
Cost of curiosity
Searching a repository and reading history is token-heavy. On a mini server with unlimited usage the Agent can afford to read widely before it writes, which is what makes its hypotheses useful.
Takeaways
- Every report becomes a structured issue.
- The Agent reads code, engineers change it.
- Unlimited usage pays for thorough reading.