Live multi-agent debate. Spawns named teammates (agent teams) that argue WITH EACH OTHER directly via SendMessage — a real, persistent debate, not one agent simulating both sides in a single reply. Heuristic: use teammates when the agents need to talk to each other (debate, collaborative brainstorm, hand-offs); use parallel subagents when they don't (independent fan-out). Keywords: debate, debate this, argue both sides, pressure-test, pressure test, steelman, devil's advocate, red team, red-team, challenge this, dialectic, adversarial review, two advocates, for and against, attack this idea. Use when: user types /debate, says 'debate this', 'argue both sides', 'pressure-test this decision', or wants a high-stakes call stress-tested by opposing positions that engage each other live. Requires the MAIN session (a subagent cannot spawn teammates) and CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Do NOT use for: simple questions, mechanical tasks, or when one analysis suffices.
Installation
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Live multi-agent debate. Spawns named teammates (agent teams) that argue WITH EACH OTHER directly via SendMessage — a real, persistent debate, not one agent simulating both sides in a single reply. Heuristic: use teammates when the agents need to talk to each other (debate, collaborative brainstorm, hand-offs); use parallel subagents when they don't (independent fan-out). Keywords: debate, debate this, argue both sides, pressure-test, pressure test, steelman, devil's advocate, red team, red-team, challenge this, dialectic, adversarial review, two advocates, for and against, attack this idea. Use when: user types /debate, says 'debate this', 'argue both sides', 'pressure-test this decision', or wants a high-stakes call stress-tested by opposing positions that engage each other live. Requires the MAIN session (a subagent cannot spawn teammates) and CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Do NOT use for: simple questions, mechanical tasks, or when one analysis suffices.
Debate
Guaranteed entry point for a live agent-team debate: named teammates argue with each other via SendMessage, you moderate, then you synthesize what survived. The real thing — not one agent role-playing both sides in a single response.
This skill is a thin trigger. The full protocol, spawn-prompt template, moderation rules, and variants live in the heavy-think skill — read it before spawning:
../heavy-think/references/team-debate.md
Prerequisites (check first)
Requirement
Why
If missing
You are the main session
Agent teams have no nested teams — a subagent can't spawn teammates
Stop; tell the user to run debate from the main session
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
Enables teammates + SendMessage
Fall back to heavy-think's Round Protocol (simulated, weaker)
If either fails, say so explicitly — do NOT silently simulate the debate and pass it off as real.
Playbook
Frame the motion. One sentence, debatable, with a clear A-vs-B or yes/no. Surface the hidden axes — what would change the verdict.
Assign stances. 2–4 positions that genuinely conflict (not shades of agreement). Name each debater.
Spawn all debaters in ONE message — model="opus", each given: full context (teammates don't inherit your history), its assigned stance, and its opponents by name. Instruct each to argue via SendMessage, engage the opponent's LATEST point, make no file edits, spawn nothing.
Opening round. Each states its strongest case.
Cross-examine. They attack each other directly, round by round. You bound rounds (max 3) and force engagement with the latest argument — never argue a stance yourself.
Detect convergence. Stop when positions stabilize or a crux is isolated.
Synthesize + shut down. YOU write the verdict: what survived challenge, what broke, the decision and its conditions. Then shut the teammates down.
Rules
Moderate, don't compete — you steer, you never take a side.
Max 3 rounds; beyond that is diminishing returns.
Anti-sycophancy: debaters attack arguments, not flatter each other.
The deliverable is YOUR synthesis, not a transcript dump.
