Tackle a research question single-agent prompts fail on
When to use: Questions that need breadth (many sources) and depth (skepticism, synthesis) — where one pass hallucinates.
Flow
-
Pose the questionUse infinite-gratitude. Research question: what are the real bottlenecks in scaling post-transformer architectures in 2026? Run the 10-agent dojo.✓ Copied→ Roles spin up: explorer, critic, synthesizer, etc., each producing a distinct contribution
-
Watch the handoffsShow me the critic's pushback on the explorer's output.✓ Copied→ Substantive critique, not agreement
-
Final synthesisGive me the synthesizer's final artifact with dissent noted.✓ Copied→ Document with main thesis + minority views preserved
Outcome: A research brief richer than a single-agent answer.
Pitfalls
- Agents converge and critic rubber-stamps — Prompt the critic explicitly to find flaws; force disagreement
Combine with: filesystem