An 8-phase multi-agent pipeline that uses structured adversarial debate to produce high-quality, well-researched outputs. Combines blind deliberation, web search diversity, confidence-weighted synthesis, and independent verification to defend against sycophancy, consensus inertia, and error cascade.
Token counts approximate (OpenAI cl100k_base encoding, within ~15% of Claude). Pricing last verified 2026-04-12.
| Phase \ Failure | Sycophancy | Majority Tyranny | Bluster Effect | Error Cascade | Context Drift | Consensus Inertia | No New Signal | Role Bleed | Complexity Creep | Decision Oscillation | Coordinator Hedging |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Phase 1: Input Preparation | ○ | ○ | ○ | ○ | ○ | ○ |
Compose clean context for this round
Three components assembled: (1) Original constraints -- verbatim, never summarized. (2) Current synthesis -- compressed into Decided / Active / Minority Report. (3) Role-specific instruction with toke…
If compressed context unavailable, fall back to last auto-saved synthesis. If none exists, start from original input.
Agents respond blind -- each critiques before proposing
Each agent responds in isolation -- no agent sees others. Web search enabled. Every agent names one weakness before proposing. Default: 2 Optimizers + Challenger + Simplifier (4 agents, matching DeepM…
2 retries per agent. Continue with N-1 if one fails (min 2). Auto-save each response. Never skip Challenger.
Challenges logged with steel man -- arbitrated by Coordinator in same call as Phase 04
Challenger objections extracted into numbered register with: steel man, objection, evidence, confidence. Relevant agents defend or concede. Phases 03 and 04 are a SINGLE Coordinator API call. The Coor…
If defense calls fail, challenge marked 'undefended' but NOT auto-adopted. Coordinator still decides.
Arbitrate challenges + synthesize (one combined call)
Single API call that handles both challenge arbitration and synthesis. Receives: all blind agent proposals, the challenge register with defenses, Simplifier flags, dimension weights, and a ROSTER NOTE…
Most critical call. 3 retries. If fails, save raw responses as fallback. Surface to user.
Score, diff, sunset, trajectory -- checks synthesis against raw proposals
One API call, structured JSON output. The Verifier sees raw agent proposals alongside the Coordinator's synthesis. If the Coordinator silently dropped an agent's high-confidence point, the raw_vs_synt…
If fails, proceed but flag verification skipped. Double-verify next round.
Decided / active / minority / original -- 800 token hard cap with triage
Compress to four sections with TRIAGE PRIORITY if 800 tokens is tight: (1) ORIGINAL CONSTRAINTS -- verbatim, never cut, always first priority. (2) DECIDED -- locked decisions, bullet points only. (3)…
If fails, pass raw synthesis. Always auto-save to storage.
Continue / fork / stop -- with oscillation detection + human override
Five checks: (1) OSCILLATION DETECTION -- hash key decisions and compare to prior rounds. (2) HUMAN REVIEW -- optional pause. (3) CONTINUE -- scores improved, active items remain. (4) FORK -- incompat…
Lightweight check, no API call. Cannot fail.
Clean deliverable with scoring context and tradeoff transparency
Separate API call. Receives synthesis + verification JSON. The formatter strips debate rhetoric but retains actionable metadata: dimension scores, known tradeoffs, constraint renegotiations, and open…
2 retries. If fails, present raw synthesis with scores appended. Save regardless.
Agents agree more each round. RLHF rewards agreement over accuracy.
If most agents agree (right or wrong), dissenters conform.
Vivid wrong beats calm right. LLMs prefer persuasive falsehoods.
Small errors compound 17x through unstructured chains.
Original constraints forgotten. Semantic mutation shifts language.
Wrong decisions solidify. Existing choices stay by default.
Same model, different personas = same knowledge, no diversity.
Shared training biases regardless of persona.
Each round adds without resistance. By round 5, unusable.
7-phase pipeline with blind deliberation, web search, challenge register
8 failure modes, 8 principles, 7 research sources
Added Simplifier role -- complexity creep was undefended
Added Phase 08 Output Formatting -- raw synthesis not usable
Mandatory critique-before-proposal for all agents
Steel man requirement for Challenger
Phases 03+04 combined into single Coordinator call -- saves 1 API call per round
| ○ |
| ○ |
| ○ |
| ○ |
| ○ |
| Phase 2: Independent Proposals + Research | ● | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ |
|---|
| Phase 3: Challenge Register + Arbitration | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ |
|---|
| Phase 4: Coordinator Synthesizes | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ |
|---|
| Phase 5: Independent Verification | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ |
|---|
| Phase 6: Compress Context + Minority Report | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ |
|---|
| Phase 7: Decision Gate + Human Review | ● | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ |
|---|
| Phase 8: Output Formatting | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ | ○ |
|---|
HOW TO USE THESE PROMPT TEMPLATES ============================================ PASTE ORDER (new Claude conversation or Project): 1. MASTER PROMPT -- paste first. Replace the three [REPLACE] sections. 2. AGENT DEFINITIONS -- paste with: "Use these agent definitions." 3. SCORING DIMENSIONS -- paste with: "Use these scoring dimensions." 4. ITERATION MODES -- paste with: "Use these iteration modes." 5. STRUCTURAL RULES -- paste with: "Embed this as CTX constant." CUSTOMIZATION: - [REPLACE: ...] markers show what to customize. - The 7 structural ROLES are universal. Keep them for any project. - The 10 structural RULES are universal. Keep as-is. - SCORING DIMENSIONS and AGENT EXPERTISE are domain-specific. Replace fully. API CALL BUDGET (revised v3.3): - Per round: ~7-9 calls - Full 3-round debate: ~25-31 total + 1 output formatter - Estimated time: 3-6 minutes per full debate - Estimated cost: ~$0.50-2.00 per debate with Sonnet
Build me a React (.jsx) multi-agent adversarial debate dashboard. Single-file default-export React component. PROJECT CONTEXT Domain: [REPLACE: describe your problem domain] Starting input: [REPLACE: paste your starting material] User priorities: [REPLACE: list quality priorities and tradeoff preferences] ARCHITECTURE -- 8-PHASE PIPELINE (v3.3) Phase 01: Input Preparation Phase 02: Independent Proposals + Research Phase 03+04: Challenge Register + Coordinator Synthesis (COMBINED CALL) Phase 05: Independent Verification Phase 06: Compress Context + Minority Report Phase 07: Decision Gate Phase 08: Output Formatting STRUCTURAL RULES: (1) Agents are BLIND (2) Must name one weakness BEFORE proposing (3) Challenger must STEEL MAN before objecting (4) Simplifier has VETO (5) Challenges REGISTERED and individually arbitrated (6) Overruled positions PRESERVED in Minority Report (7) Coordinator evaluates by confidence + evidence, not rhetoric (8) Word limits enforced per agent (9) Be specific: exact numbers, measurements, specifications (10) Max 3 rounds. Auto-stop on score delta < 0.5 (11) CONSTRAINT RENEGOTIATION: formally flag, never silently drop
Round N undoes Round N-1. Binding rules flip decisions.
Averages positions instead of deciding. Non-committal.
Verifier independence: now receives raw agent proposals alongside synthesis
DIMS scale fields injected into Verifier prompt at build time -- scoring anchored consistently across rounds