A bug comes in from users. Your program crashes with an error at some indeterminate moment, for unclear reasons. (Are there ever any other kind?)
It’s serious. Lots of users complaining. It’s urgent — every engineer drops what they’re doing and piles onto the task. What’s the CPU doing? Which processes are running? Are there logs? Are there repro steps? What about the dump? Etc.
The war room
Only they’re not exchanging files anymore. They’re exchanging prompts.
- The first analyzed the logs with Claude.
- The second asked Codex to reproduce it.
- The third ran DeepSeek over all of those prompts and assembled an uber-prompt.
- The fourth fed the repro steps into Gemini and spins them in a loop, catching crashes.
- The fifth is having the commits from that time window analyzed.
The task has a ton of comments. Everyone is busy. Everyone is contributing. Prompt profiling. Prompt performance engineering.
The part nobody mentions
Each of them has a gut feeling that something is wrong in exactly this chunk of code.
And the agent completely agrees with them. Every time. With all five of them. About five different chunks of code.
The manager looks at the activity on the task — the comment count, the velocity, the sheer momentum of it all — and honestly? Things did get better.
Better with AI.