Writing
Multi-agent software development
More agents are not more leverage unless the interfaces are boring and the tests are real.
Multi-agent setups fail in the same way distributed systems fail: unclear ownership, duplicated work, and no single place where truth is checked.
When another agent is worth it
- A role with a different tool set (retrieve vs. patch vs. run tests).
- A long-running job that should not sit in the interactive loop.
- A critic that only has permission to run the harness, not to edit it.
When it is costume
- Agents that “debate” without a shared fixture.
- Extra coordinators that restate the prompt.
- Parallel editors on the same file with no merge story.
Prefer simple orchestration. If you cannot draw the data flow on one line, you probably added a meeting.