67
One attempt is a sample size of one. This runs six, each forced down a different road, then a seventh agent judges the field, ranks it, and builds the final from the best parts. You ship the strongest version that exists, not the first one that came out.
When to use
- High-stakes deliverables: the pitch, the architecture, the pricing page, the cold email.
- Open-ended problems with many valid approaches and no obvious winner.
- Not for trivial tasks. Six agents on a one-liner is theater.
Behavior
1. Frame once
Write a single task brief every agent receives verbatim: the task, the deliverable format, and the constraints. Tight enough that all six outputs are directly comparable.
2. Fan out six agents in parallel
Spawn all six in one message (Agent/Task tool). Each gets the same brief plus ONE strategy lock it must follow:
| # | Lens | Instruction |
|---|---|---|
| 1 | Minimalist | Smallest thing that fully works. Cut everything optional. |
| 2 | Maximalist | Most complete version. Cover every edge and every objection. |
| 3 | First principles | Ignore convention. Derive the answer from scratch. |
| 4 | Prior art | Find how the best existing versions do it and adapt the strongest pattern. |
| 5 | Adversarial | Design against failure. Ask what breaks each choice, keep what survives. |
| 6 | Wildcard | One unreasonable creative bet, executed seriously. |
Each agent must return a complete, standalone answer. No hedging, no "it depends" without a decision.
If no sub-agent tool is available, run six sequential passes with a hard reset between each: "discard the previous approach entirely; you are now agent N with this lens."
3. Agent seven judges
The judge never writes its own answer first. It scores each attempt 1 to 10 on correctness, completeness, and taste; ranks all six with a one-line verdict each; and names each attempt's single strongest element, including the losers'.
4. Synthesize
Build the final on the winner's skeleton, then graft in every strongest element worth stealing. Note what was taken from whom.
Output
The leaderboard (rank, score, one-line verdict per agent), then the synthesized final, then a short "stolen parts" ledger.
