# Multi Round Deliberation > Reusable protocol for multi-agent discussions that iterate until convergence. Used by results-integration (Phase 5) and paper-writing (Phase 6) to ensure discussions actually resolve issues rather than just noting them. - Skill: `evoclaw/multi-round-deliberation` (Agent Skill) - Install (CLI): `npx skillmds@latest add evoclaw/multi-round-deliberation` - Raw SKILL.md: https://api.skillmd.com/api/skills/evoclaw/multi-round-deliberation/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: EvoClaw (https://skillmd.com/u/evoclaw) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/evoclaw/multi-round-deliberation --- # Multi-Round Deliberation Protocol ## Overview A single round of feedback is NOT a discussion. Real deliberation requires iteration: agents identify issues → someone fixes them → agents verify the fix → repeat until resolved. This protocol standardizes this loop across all phases. ## The Problem with Single-Round Feedback ``` WRONG (current anti-pattern): 3 agents give opinions → main agent "synthesizes" → done → Issues may be noted but never actually resolved → No verification that fixes addressed the concerns RIGHT (this protocol): 3 agents give opinions → main agent modifies artifact → agents re-check → if unresolved, modify again → repeat until convergence or max rounds ``` ## Core Protocol ### Shared Value Framework (prevents divergence) All agents in a deliberation share these values, injected into every prompt: ``` SHARED VALUES (inject into every agent prompt in this deliberation): ═══════════════════════════════════════════════════════════════════ Target venue: [from research-anchor.yaml] Value proposition: [from research-anchor.yaml] Research type: [from research-anchor.yaml] OPTIMIZATION TARGET: "Would this survive peer review at [target venue]?" SCORING RUBRIC (default — phases may override; see phase-specific sections): - PASS: No fatal or major issues. Ready to proceed. - CONDITIONAL: Major issues exist but are addressable. Needs another round. - FAIL: Fatal issues. Fundamental rethinking needed. CONVERGENCE RULE: Deliberation ends when ALL agents score PASS, or when max rounds are reached and remaining issues are presented to user. ANTI-DIVERGENCE: If you disagree with another agent's feedback, state specifically WHY and propose a CONCRETE alternative. "I disagree" without a counter-proposal is not constructive. ═══════════════════════════════════════════════════════════════════ ``` ### Deliberation Loop ``` ┌─────────────────────────────────────────────────────┐ │ ROUND N │ │ │ │ 1. ASSESS (parallel) │ │ Dispatch ALL agents with current artifact │ │ Each returns: issues + verdict (PASS/COND/FAIL) │ │ │ │ 2. CHECK CONVERGENCE │ │ All PASS? → END (consensus reached) │ │ Any FAIL or COND + round < max? → go to step 3 │ │ Max rounds (5) reached? → END (present to user) │ │ │ │ 3. MODIFY │ │ Main agent (you) incorporates feedback: │ │ - Fix all issues marked as addressable │ │ - For disagreements: choose the stronger argument│ │ - Update the artifact │ │ │ │ 4. FULL RE-ASSESS (next round) │ │ Dispatch ALL agents again with: │ │ - The complete modified artifact │ │ - Summary of changes made since last round │ │ - Previous round's issues for reference │ │ All agents review the full artifact — not just │ │ their own previous concerns — because │ │ modifications can introduce new issues. │ │ │ │ → Back to step 2 │ └─────────────────────────────────────────────────────┘ ``` ### Round Limits (non-negotiable) | Context | Max Rounds | Rationale | |---------|-----------|-----------| | Phase 5 story design | 5 | High-stakes — the story determines the paper's quality | | Phase 6 per-section polishing | 5 | Each section must be thoroughly vetted | | Phase 6 full-paper review | 5 | Full-paper coherence is critical | ### Convergence Criteria Deliberation ends (consensus) when ALL of: - No agent gives verdict "FAIL" - No remaining issues tagged "fatal" or "critical" - All agents explicitly state their concerns from previous rounds are addressed ### Non-Convergence Handling If max rounds reached and agents still disagree: ``` DELIBERATION SUMMARY (presented to user): ══════════════════════════════════════════ Rounds completed: N / max N Final verdicts: Agent A: [PASS/COND/FAIL], Agent B: [...], Agent C: [...] RESOLVED issues (N): ✅ [issue] — addressed in round [N] by [change] UNRESOLVED issues (N): ⚠️ [issue] — Agent [X] says: "[position]" — Agent [Y] says: "[counter-position]" — My recommendation: [your judgment] DECISION NEEDED from you: 1. [Option A — side with Agent X] 2. [Option B — side with Agent Y] 3. [Option C — compromise you propose] 4. Run one more round of discussion ``` ### Full Re-Assessment Every Round Every round, dispatch ALL agents with the complete modified artifact. Do NOT skip any agent — modifications to address one agent's concerns can introduce new issues that another agent would catch. **Re-assessment prompt template (used for Round 2+):** ``` Call Task tool with: description: "[agent role] — round [N] review" prompt: | [SHARED VALUES block] This is round [N] of deliberation. In the previous round, you and other agents raised these concerns: === YOUR previous concerns: [paste this agent's issues from last round] OTHER agents' concerns (for context): [paste summary of other agents' issues] === Changes made since last round: === [paste summary of what changed and why] === COMPLETE MODIFIED ARTIFACT: [paste the full updated artifact] Review the COMPLETE artifact (not just the changes). Modifications can introduce new issues. Answer: 1. For EACH of your previous concerns: RESOLVED / PARTIALLY / NOT ADDRESSED 2. Any NEW issues introduced by the modifications? 3. Any issues with how OTHER agents' concerns were addressed? 4. Overall verdict: PASS / CONDITIONAL / FAIL If CONDITIONAL or FAIL: state exactly what remains to be fixed. subagent_type: "generalPurpose" ``` ### Agent Disagreements When agents directly contradict each other (e.g., Story Architect says "lead with stability" but Audience Specialist says "lead with monocyte substructure"): 1. **Present both positions** with their reasoning 2. **Evaluate evidence**: which position is better supported by the actual results and venue expectations? 3. **Choose the stronger argument** for the modification 4. **In re-assessment**: tell the overruled agent what you chose and why — they may accept or push back with new arguments 5. **If persistent disagreement after 2 rounds**: present to user as a strategic choice ### Anti-Patterns to Avoid | Anti-pattern | Why it's wrong | What to do instead | |-------------|---------------|-------------------| | Accepting "CONDITIONAL" without asking what's needed | "Conditional" means there ARE remaining issues | Ask the agent to specify exactly what remains | | Running max rounds mechanically even when consensus reached in round 1 | Wastes time and tokens | Check convergence after every round; stop early if PASS | | Ignoring one agent's feedback because the other two disagree | Minority opinion may be right | Address the concern or explain why it's overruled | | "Synthesizing" by averaging opinions | Synthesis is not averaging | Choose the strongest argument for each issue | | Skipping agents in later rounds to "save cost" | Fixes can introduce new issues that other agents would catch | Dispatch ALL agents every round | ## Integration with Phases ### Phase 1 — Idea Brainstorming ``` Artifact being deliberated: 5 candidate research ideas (idea cards from Step 5c) Agents: Visionary Researcher, Pragmatic Advisor, Field Scout Max rounds: 5 SCORING RUBRIC (Phase 1 override — ideas are not pass/fail): - STRONG: High-potential idea. Clearly novel, feasible, and well-positioned. - VIABLE: Promising but needs refinement. Addressable concerns exist. - WEAK: Significant issues (novelty, feasibility, or positioning). Salvageable only if combined with another idea or substantially reworked. - KILL: Fatal flaw (already done, infeasible, or no clear contribution). Remove. Mapping to core protocol: STRONG = PASS, VIABLE = CONDITIONAL, WEAK/KILL = FAIL. Convergence uses Phase 1 rubric (STRONG/VIABLE), not core rubric (PASS/CONDITIONAL). Convergence: Agents agree on a ranking of top 2-3 ideas (all rated STRONG or VIABLE) Modifier: Main agent merges, splits, refines, or eliminates ideas between rounds Re-assessment: ALL 3 agents review complete modified idea set every round Special rules: - Ideas can die, merge, or be born during brainstorming - Visionary can propose new combinations of existing ideas - If ALL ideas rated WEAK or KILL by majority → trigger "Idea Rescue" (see below) - On-demand literature search allowed mid-deliberation (e.g., Scout needs to check competition) Idea Rescue (when all ideas fail): 1. Agents identify WHY all ideas failed — common root cause? 2. Trigger broadened literature search targeting the identified gap 3. Main agent generates 3 new idea cards based on rescue search results 4. Resume deliberation with new ideas (round counter does NOT reset) 5. If still no STRONG/VIABLE after max rounds → present situation to user with explanation of what was tried and why it failed. User decides next steps. ``` ### Phase 2 — Research Question Deliberation ``` Artifact being deliberated: Research Question Type M/D/H agents: Senior Professor, Journal/Conference Editor, Research Scientist (PhD/Postdoc) Type C agents: Target User Representative, Journal/Conference Editor, Software Architect Max rounds: 5 Convergence: All 3 give PASS on the research question Modifier: Main agent refines question between rounds Re-assessment: ALL 3 agents review complete refined question every round Special rule (Type M/D/H): Professor "FAIL — not novel" persistent for 2+ rounds → present to user immediately Special rule (Type C): Target User "FAIL — no real pain point" persistent for 2+ rounds → present to user immediately ``` ### Phase 3 — Method/Framework Design Deliberation ``` Type M artifact: Method design (innovation point, architecture, baselines, ablations) Type M agents: Innovation Advisor, Technical Architect, Baseline Devil's Advocate Type D artifact: Analysis framework (analysis plan, sufficiency criteria, alternatives) Type D agents: Domain Scientist, Methodology Consultant, Statistical Rigor Advisor Type C artifact: Tool design (utility claim, evaluation framework, competing tools, case studies) Type C agents: Target User, Competing Tool Expert, Software Quality Advisor Max rounds: 5 Convergence: All 3 give PASS on the design Modifier: Main agent refines design between rounds (may search additional literature) Re-assessment: ALL 3 agents review complete refined design every round Special rule: If any agent flags missing literature → search before next round ``` ### Phase 5 — Story Design Deliberation ``` Artifact being deliberated: Argument Blueprint Agents: Story Architect, Devil's Advocate, Audience Specialist Max rounds: 5 Convergence: All 3 give PASS on the blueprint Modifier: Main agent rewrites blueprint between rounds Re-assessment: ALL 3 agents review complete blueprint every round Special rule: Devil's Advocate "fatal" on novelty → HARD BLOCK (no amount of rounds fixes missing novelty) ``` ### Phase 6 Per-Section — Writing Polishing Deliberation ``` Artifact being deliberated: One LaTeX section Agents: Domain Expert, Writing Editor, Adversarial Reviewer Max rounds: 5 Convergence: All 3 give PASS; authority score ≥ 7 Modifier: Main agent rewrites section between rounds Re-assessment: ALL 3 agents review complete section every round Special rule: If Adversarial Reviewer verdict "NO" after final round → trigger Cross-Critique (Phase B2) ``` ### Phase 6 Full-Paper — Final Review Deliberation ``` Artifact being deliberated: Complete paper Agents: Proofreader, Domain Expert, Adversarial Reviewer #2 Max rounds: 5 Convergence: No FATAL/CRITICAL issues remaining Modifier: Main agent fixes specific sections between rounds Re-assessment: ALL 3 agents review complete paper every round Special rule: "EXPERIMENT NEEDED" from any agent → return to Phase 4 (not fixable by more rounds) ```