Multi-Agent: Fan-Out, Adversary, Synthesis

There's a second flavor of "many agents" that runs within a single task rather than across separate ones. Instead of one agent doing everything, you split the roles:

• Fan-out for breadth. Several agents tackle the same problem from different angles — or explore different parts of a large codebase — then report back. Useful when you want coverage: "find every place we call this deprecated function" parallelizes well.
• An adversarial reviewer. One agent writes; a second agent's only job is to attack the first one's output — find the bug, the missing edge case, the unhandled error. A reviewer pointed at a diff with the instruction "assume this is wrong and find out why" catches things the author agent talked itself into.
• Synthesis. A final pass collects the findings and produces one coherent result — the working code plus a short note on what the reviewer flagged and how it was resolved.

Reviewer agent brief:

You are reviewing a diff written by another agent. Assume it contains
at least one bug and find it. Specifically check:
- Does it actually satisfy the original goal, or just pass the tests?
- Edge cases: empty input, nulls, the largest realistic input.
- Anything deleted or changed that the task didn't ask for.

Report concrete problems with file and line. Do not rewrite the code —
just report. If you find nothing after a genuine look, say so plainly.

The adversarial reviewer is worth setting up even when you're one person with no fleet of agents — because the author of a change is the worst person to find its flaws, human or AI. A second pass with a hostile prompt is cheap insurance.