Quick fixes AG-O001
Add a critic, not a committee
Shared methods · A shared method; linked tool guides explain the exact steps.
This page covers tools outside your selection. You can still read it. Find matching guides
One independent critic measurably improves agent work. A panel of role-play agents mostly agrees with itself, at several times the cost.
My agent review panel agrees with itself? Here's the quick fix ↓ — the why sits right below it.
Quick fix: Add a critic, not a committee
Instead of Spawn a five-agent review board: architect persona, security persona, style persona, product persona, and a judge.
Try Spawn one reviewer with fresh context and one instruction: find what is wrong with this, and try to refute its claims. Add a second only with a genuinely different lens.
Why this works
Agents asked to reach agreement tend to agree, with the work and with each other. Adding more of them multiplies cost and adds correlated opinions, so the failure the panel was meant to catch survives a unanimous vote. What breaks the pattern is adversarial framing, and small numbers: a reviewer plus a critic instructed to disagree exposes weak reasoning that a consensus-shaped committee papers over. The empirical version: a minimal two-agent reviewer-critic loop outscored both a single reviewer and a four-role panel on code benchmarks. Structure of dissent beats headcount.
When this applies
You are about to give agents roles (names, personas, a meeting) rather than oppositions (one builds, one attacks), or a review step returns polite agreement on work you suspect is flawed.
Template
You are reviewing work you did not produce. Your job is to refute it: find the strongest case that it is wrong, unsafe, or incomplete. If you cannot, say what you checked. Do not summarise it; attack it.
Go deeper
The mechanics of independent review contexts are in subagents, and why a fresh-context critic sees what the author-session cannot is the same argument as reviewing agent code. Model agreement is weak evidence in general — a second model is weaker corroboration than it appears.
Improve next
- Ask for three options, not one answer
Ask for two or three options before committing to the first.
- Don't fan out past your review capacity
You can generate more than you can review — the reading is the limit.
Something wrong with this page?
Say what you expected and what you got. That is usually the shortest route to a correction, and it goes on the public issue tracker so the fix is visible.