build-consensus
について
このスキルは、閾値投票やクォーラムセンシングなどのメカニズムを用いて、中央権限を介さない分散合意を実現します。提案の生成、支持動態、リーダー不在の集団意思決定におけるデッドロック解消を扱います。中央集権的な意思決定がボトルネックとなる場合や、分散データベースやマルチエージェントAIのような自動化された合意形成システムを設計する際にご活用ください。
クイックインストール
Claude Code
推奨npx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/build-consensusこのコマンドをClaude Codeにコピー&ペーストしてスキルをインストールします
ドキュメント
Build Consensus
Achieve collective agreement across distributed agents without central authority — use scout advocacy, threshold quorum sensing, commitment dynamics modeled on honeybee swarm decision-making.
When Use
- Group must collectively decide between multiple options without designated leader
- Centralized decision-making is bottleneck or single point of failure
- Stakeholders have different information, perspectives that must be integrated
- Past decisions suffered from groupthink (premature convergence) or analysis paralysis (no convergence)
- Designing automated systems that must reach consensus (distributed databases, multi-agent AI)
- Complementing
coordinate-swarmwhen coordination requires explicit collective decisions
Inputs
- Required: Decision to be made (binary choice, selection from N options, parameter setting)
- Required: Participating agents (team members, services, voters)
- Optional: Known options with preliminary quality assessments
- Optional: Decision urgency (time budget)
- Optional: Acceptable error rate (can group occasionally pick second-best option?)
- Optional: Current decision-making failure mode (groupthink, deadlock, flip-flopping)
Steps
Step 1: Generate Proposals Through Independent Scouting
Ensure decision space adequately explored before any advocacy begins.
- Assign scouts to independently explore option space:
- Each scout evaluates options without knowing other scouts' findings
- Independent evaluation prevents early herding toward popular-but-mediocre options
- Scout count: minimum 3 scouts per serious option (for reliability)
- Scouts produce structured assessments:
- Option identifier
- Quality score (normalized 0-100 or categorical: poor/fair/good/excellent)
- Key strengths and risks identified
- Confidence level (how thoroughly was option evaluated?)
- Aggregate scout reports without filtering — all options above minimum quality threshold enter advocacy phase
Got: Set of independently evaluated proposals with quality scores and assessments. No option eliminated by single evaluator; diversity of perspective preserved.
If fail: Scouts converge on same option without independent evaluation? Scouting not truly independent. Rerun with explicit information barriers. Too many options survive to advocacy phase? Raise minimum quality threshold. Too few survive? Lower it or add more scouts.
Step 2: Run Advocacy Dynamics (Waggle Dance)
Allow scouts to advocate for preferred options. Advocacy intensity proportional to quality.
- Each scout advocates for their top-rated option:
- Advocacy intensity proportional to quality score (better options get more vigorous advocacy)
- Advocacy public — all agents observe all advocacy signals
- Advocates present evidence and quality assessment, not just preference
- Uncommitted agents observe advocacy and evaluate:
- Follow up on advocated options by inspecting them independently
- Agent's own inspection confirms quality → joins advocacy
- Inspection reveals lower quality than advertised → does not join
- Cross-inspection dynamics:
- Advocates for weaker options naturally lose followers as agents independently verify
- Advocates for stronger options gain followers through confirmed quality
- Process self-correcting: exaggerated advocacy fails verification step
Advocacy Dynamics:
┌─────────────────────────────────────────────────────────┐
│ Scout A advocates Option 1 (quality 85) ──→ ◉◉◉◉◉ │
│ Scout B advocates Option 2 (quality 70) ──→ ◉◉◉ │
│ Scout C advocates Option 3 (quality 45) ──→ ◉ │
│ │
│ Uncommitted agents inspect: │
│ Agent D inspects Option 1 → confirms → joins ◉◉◉◉◉◉ │
│ Agent E inspects Option 2 → confirms → joins ◉◉◉◉ │
│ Agent F inspects Option 3 → disagrees → inspects Opt 1│
│ → confirms → joins ◉◉◉◉◉◉◉│
│ │
│ Over time: Option 1 advocacy grows, Option 3 fades │
└─────────────────────────────────────────────────────────┘
Got: Advocacy for best option(s) grows over time as agents independently verify quality. Advocacy for weaker options fades as verification fails. Group naturally converges toward strongest option without any agent dictating choice.
If fail: Advocacy doesn't converge (two options remain neck-and-neck)? Options may be genuinely equivalent — proceed to quorum with either, or use tiebreaker rule. Advocacy converges too fast on mediocre option? Increase independence of evaluation (more scouts, stricter information barriers), add mandatory cross-inspection step.
Step 3: Set Quorum Threshold and Commit
Define commitment threshold that triggers collective action.
- Set quorum threshold:
- Simple decisions: 50% + 1 of agents committed to one option
- Important decisions: 66-75% committed to one option
- Critical/irreversible decisions: 80%+ committed to one option
- Rule of thumb: higher stakes → higher quorum → slower but more reliable consensus
- Monitor commitment accumulation:
- Track how many agents have committed to each option over time
- Display commitment levels transparently (all agents can see current state)
- Do not allow commitment withdrawal mid-cycle (prevents oscillation)
- Quorum reached:
- Winning option adopted as collective decision
- Advocates for losing options acknowledge decision (no rogue agents)
- Implementation begins immediately — delay after consensus erodes commitment
Got: Clear quorum moment where enough agents have independently committed to one option. Decision legitimate because emerged from independent evaluation, not authority or coercion.
If fail: Quorum never reached within time budget? Escalate to Step 4 (deadlock resolution). Quorum reached but agents unhappy? Advocacy phase was too short — agents committed without adequate evaluation. Consensus was wrong (discovered after fact)? Independent scouting insufficient — increase scout diversity and evaluation thoroughness in next cycle.
Step 4: Resolve Deadlocks
Break decision gridlock when natural consensus process stalls.
- Diagnose deadlock type:
- Genuine tie: two options equally good → flip coin; cost of delay exceeds cost of picking "wrong" equal option
- Information deficit: agents can't evaluate options well enough → invest in more scouting before re-running advocacy
- Faction formation: entrenched subgroups refuse to cross-inspect → introduce mandatory rotation where advocates must inspect opposing option
- Option proliferation: too many options fragment commitment → eliminate bottom 50%, re-run advocacy
- Apply appropriate resolution:
- Genuine tie: random selection or merge options if compatible
- Information deficit: time-boxed scouting extension
- Faction formation: forced cross-inspection round
- Option proliferation: ranked elimination tournament
- After resolution, reset quorum clock, re-run Step 3
Got: Deadlock resolved through appropriate intervention. Resolution visible and accepted by group as fair process, even if individual agents preferred different outcome.
If fail: Deadlocks recur on same decision? Decision framing may be wrong. Step back, ask: can decision be decomposed into smaller, independent decisions? Can scope be reduced? Is there "try both and see" option? Sometimes best consensus is "we'll run time-boxed experiment."
Step 5: Assess Consensus Quality
Evaluate whether consensus process produced good decision, not just decision.
- Post-decision assessment:
- Was winning option independently verified by at least N agents?
- Was decision speed appropriate (not too fast/groupthink, not too slow/paralysis)?
- Did process surface information that would have been missed by single decision-maker?
- Are agents committed to implementation, or merely compliant?
- Track consensus health metrics:
- Time to quorum: decreasing over successive decisions indicates learning; increasing indicates growing complexity or dysfunction
- Scout-to-commit ratio: how much scouting was needed per commitment? High ratio = difficult decision or low trust
- Post-decision regret rate: how often does group wish it had chosen differently?
- Feed learnings back into process:
- Adjust quorum thresholds based on decision importance and past accuracy
- Adjust scout count based on option complexity
- Adjust time budgets based on historical time-to-quorum
Got: Feedback loop that improves consensus quality over time. Group learns to scout more effectively, advocate more honestly, commit more confidently.
If fail: Consensus quality metrics poor (high regret, slow decisions)? Audit process for structural failures: insufficient scouting diversity, advocacy without verification, or thresholds set too low for decision type. Rebuild specific failing stage rather than overhauling entire process.
Checks
- Proposals generated through independent scouting (no herding)
- Advocacy intensity proportional to assessed quality
- Uncommitted agents independently verified advocated options
- Quorum threshold appropriate for decision's importance
- Quorum reached, decision implemented promptly
- Deadlock resolution mechanism available (even if unused)
- Post-decision quality assessment conducted
Pitfalls
- Skipping independent scouting: Jumping directly to advocacy produces groupthink. Quality of consensus depends entirely on quality of independent evaluation
- Equal advocacy for unequal options: Every option gets same advocacy regardless of quality? Process degenerates into random selection. Advocacy must be proportional to assessed quality
- Commitment withdrawal: Allowing agents to un-commit creates oscillation. Once committed in cycle, agents stay committed until cycle resolves
- Confusing consensus with unanimity: Consensus requires sufficient agreement, not total agreement. Waiting for 100% creates permanent deadlock
- Ignoring losing side: Agents who advocated for losing option have information group needs. Their concerns should inform implementation, even if they don't block decision
See Also
coordinate-swarm— foundational coordination framework supporting signal-based consensus mechanismdefend-colony— collective defense decisions often require rapid consensus under threatscale-colony— consensus mechanisms must adapt when group size changes significantlydissolve-form— morphic skill for controlled dismantling, where consensus before dissolution is criticalplan-sprint— sprint planning involves team consensus on commitment scopeconduct-retrospective— retrospectives are form of consensus-building about process improvementbuild-coherence— AI self-application variant; maps bee democracy to single-agent multi-path reasoning with confidence thresholds and deadlock resolution
GitHub リポジトリ
関連スキル
content-collections
メタこのスキルは、Content Collections(Markdown/MDXファイルを型安全なデータコレクションに変換するTypeScriptファーストのツール)の本番環境でテストされた設定を提供します。Zodバリデーションによる型安全性を実現し、ブログ、ドキュメントサイト、コンテンツ重視のVite + Reactアプリケーション構築時にご利用ください。Viteプラグインの設定、MDXコンパイルから、デプロイ最適化、スキーマバリデーションまで、すべてを網羅しています。
polymarket
メタこのスキルは、開発者がPolymarket予測市場プラットフォームを活用したアプリケーション構築を可能にします。API統合による取引や市場データの取得に加え、WebSocketを介したリアルタイムデータストリーミングにより、ライブ取引や市場活動を監視できます。取引戦略の実装や、ライブ市場更新を処理するツールの作成にご利用ください。
creating-opencode-plugins
メタこのスキルは、開発者がコマンド、ファイル、LSP操作など25種類以上のイベントタイプにフックするOpenCodeプラグインを作成することを支援します。JavaScript/TypeScriptモジュール向けに、プラグイン構造、イベントAPI仕様、および実装パターンを提供します。カスタムイベント駆動ロジックでOpenCode AIアシスタントのライフサイクルをインターセプト、監視、または拡張する必要がある場合にご利用ください。
sglang
メタSGLangは、高性能なLLMサービングフレームワークであり、RadixAttentionプレフィックスキャッシュを活用したJSON、正規表現、エージェントワークフロー向けの高速で構造化された生成を特長とします。特にプレフィックスが繰り返されるタスクにおいて、大幅に高速な推論を実現し、複雑な構造化出力やマルチターン対話に最適です。制約付きデコードが必要な場合や、広範なプレフィックス共有を伴うアプリケーションを構築する場合は、vLLMなどの代替案ではなくSGLangを選択してください。
