multi-agent-refine

name

multi-agent-refine

description

Take a piece of work and make it genuinely solid by running it through several critic subagents from different angles, iterating until it holds up. Works on anything - a PRD, plan, strategy, email, piece of copy, or a decision. Does a quick brief-check first so the critics judge against the right bar, spins up task-adaptive critics in parallel, and loops (max 3 rounds, stopping early when a round is clean). Asks up front whether to refine autonomously or with your input each round. Use when the user says "refine this", "multi-agent-refine", "pressure-test this", "make this solid", "tear this apart", or is about to finalise a substantial deliverable.

/multi-agent-refine

Make any piece of work solid by putting it under adversarial pressure from several angles at once, repeated until it survives. Not one polish pass - multiple critics genuinely trying to break it, then fixing what they find, then doing it again.

Severity, used throughout (so "solid" is executable):

• HIGH = would actually fail, mislead, or harm if shipped as-is.
• MED = a real weakness worth fixing.
• LOW = nitpick / taste. Don't spend rounds on these. A round is clean when it surfaces no new HIGH or MED. That's the stop signal.

Step 0 - Quick brief-check + mode

1. Ask 1-2 quick questions so the critics judge against the right bar: what is this for and who's it for, and what would make it "solid"? If the user can't articulate the bar, don't stall - infer a sensible one, state it in one line, and proceed. (Building from scratch rather than handing you a draft? Interview a bit more first.)
2. One-line preview, then ask the mode: "I'll run ~4 critics from different angles, fix what they find, and repeat up to 3 times. Want me to do that autonomously, or check in with you each round?" If they choose input, use AskUserQuestion between rounds.

Step 1 - Pick the critics (task-adaptive, genuinely distinct)

Choose the lenses most relevant to THIS work - 3-4 critics (never more than 5, and never run more than ~8 calls total across all rounds, to stay under Cowork's ~10 cap). Examples:

• PRD / plan / spec: feasibility, scope & over-build, what's-missing, devil's-advocate
• copy / email: clarity, voice, what-a-sceptic-reads-into-it
• strategy / decision: strongest counter-argument, second-order effects, "what makes this fail"

Make the lenses genuinely different from each other - these are the same model wearing different hats, so distinct, specific framings (not five flavours of "is it good?") are what produce real coverage. Always include a "what's missing / what would a sharp critic say" lens. Only if the work rests on external facts (tech behaviour, prices, claims), add a fact-check critic (Perplexity/Firecrawl).

Step 2 - Run the critics in parallel (explicit instruction)

Invoke them explicitly - Cowork only runs subagents in parallel when told to. Say, in effect: "Spin up N subagents in parallel; give each one the work, the bar from Step 0, and ONE lens; have each return its findings." Tell each critic to genuinely try to break the work and default to sceptical - a critic that says "looks good" without hunting has failed. Each returns: concrete problems tagged HIGH/MED/LOW, and "solid on my axis: yes/no".

Step 3 - Synthesise (keep every real issue)

Merge the critiques: dedupe, and keep every HIGH and MED - only drop LOW nitpicks. Lead with anything two+ critics agreed on. Don't quietly bury a HIGH because it's inconvenient to fix.

• Autonomous mode: fix the HIGH/MED issues, show a short "what I changed and why".
• Input mode: show the findings, ask (AskUserQuestion) which to act on, then fix.

Step 4 - Loop until solid (max 3 rounds, cost-aware)

Re-critique the revised work - but only re-check the issues that were fixed plus anything new the changes introduced, with fewer critics (1-2) on later rounds, not a full fresh panel. Stop when a round is clean (no new HIGH/MED), or after 3 rounds. If round 1 is already clean, say "this is already solid" and stop - don't manufacture issues. If it's still not solid at the cap, say so plainly and name what's still weak.

Step 5 - Close

Hand back the refined work plus one line on what got stronger and anything still worth watching.

Notes

• Honest cost: this spends real tokens (several subagents per round). Keep it for things worth getting right; the brief-check + 3-4 critics + early-exit keep it bounded.
• Same-model caveat: the critics are one model in different hats, so this is perspective-diversity, not true independence. Genuinely distinct, sceptical lenses are what make it useful rather than theatre.