スキル一覧に戻る

scale-colony

pjt222
更新日 2 days ago
2 閲覧
17
2
17
GitHubで表示
その他ai

について

このスキルは、出芽や役割分化などのメカニズムを用いて、分散システムとチームを生物学的コロニーに模倣することで、スケーリング戦略を提供します。成長段階を認識し、規模拡大に伴う調整不全を防ぐためのアーキテクチャ移行を実装するのに役立ちます。コミュニケーションのオーバーヘッドが生産的な成果を上回った場合や、小規模では機能していたシステムが成長に伴って破綻する場合に活用してください。

クイックインストール

Claude Code

推奨
メイン
npx skills add pjt222/agent-almanac -a claude-code
プラグインコマンド代替
/plugin add https://github.com/pjt222/agent-almanac
Git クローン代替
git clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/scale-colony

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

Scale Colony

Scale distributed sys|teams|orgs → budding (split), role diff (age polyethism), growth-triggered arch transitions — maintain coord quality as colony grows beyond initial design.

Use When

  • Worked @ 10 agents, breaks @ 50
  • Comms overhead > productive output
  • Implicit coord patterns need explicit
  • Plan growth → scale proactive
  • Coord fails correlate w/ size (lost msgs, dup work, unclear ownership)
  • Existing sys needs split → semi-autonomous sub-colonies

In

  • Required: Current size + target growth
  • Required: Current coord mechanisms + stress points
  • Optional: Structure (flat|hierarchical|clustered)
  • Optional: Role diff already in place
  • Optional: Growth timeline + constraints
  • Optional: Inter-colony coord needs (if splitting)

Do

Step 1: Recognize Growth Phase

Identify scaling phase → apply right strategy.

  1. Classify phase:
Colony Growth Phases:
┌───────────┬──────────────┬───────────────────────────────────────────┐
│ Phase     │ Size Range   │ Characteristics                           │
├───────────┼──────────────┼───────────────────────────────────────────┤
│ Founding  │ 1-7 agents   │ Everyone does everything, direct comms,   │
│           │              │ implicit coordination, high agility       │
├───────────┼──────────────┼───────────────────────────────────────────┤
│ Growth    │ 8-30 agents  │ Roles emerge, some specialization, comms  │
│           │              │ overhead increases, need for structure     │
├───────────┼──────────────┼───────────────────────────────────────────┤
│ Maturity  │ 30-100 agents│ Formal roles, layered coordination,       │
│           │              │ sub-groups form, inter-group coordination  │
├───────────┼──────────────┼───────────────────────────────────────────┤
│ Fission   │ 100+ agents  │ Colony too large for single coordination  │
│           │              │ framework, must bud into sub-colonies     │
└───────────┴──────────────┴───────────────────────────────────────────┘
  1. Stress signals:
    • Comms overload: msgs/agent/day grows faster than colony size
    • Decision latency: proposal→decision time ↑
    • Coord failures: dup work, dropped tasks, conflicting actions ↑
    • Knowledge dilution: newcomers slow to productive
    • Identity loss: agents can't describe purpose consistently
  2. About to cross phase boundary or already crossed?

→ Clear phase ID + stress signals indicating approach|cross.

If err: phase unclear → measure 3 metrics: comm vol/agent, decision latency, coord fail rate. Plot over time. Inflection points = phase transitions. No metrics → likely Founding (where metrics not yet needed).

Step 2: Role Differentiation (Age Polyethism)

Progressive specialization → roles by experience + colony needs.

  1. Role progression:
    • Newcomers: observation, learning, simple (low autonomy, high guidance)
    • Workers: standard exec, signal following (mod autonomy)
    • Specialists: domain expertise, complex tasks, mentor newcomers (high autonomy)
    • Foragers/Scouts: exploration, innovation, external interface (see forage-resources)
    • Coordinators: inter-group comms, conflict resolution, quorum mgmt
  2. Role transitions:
    • Triggered by experience thresholds, not appointment
    • Agent done threshold tasks successfully → next role (calibrate by complexity + growth rate — 5-10 simple, 20-30 specialist)
    • Reverse possible (specialist → worker in new domain)
    • Distribution adapts to needs:
      • Growing → more newcomer slots, active mentoring
      • Stable → balanced across all roles
      • Threatened → more defenders, fewer scouts (see defend-colony)
  3. Preserve flexibility:
    • No agent permanently locked
    • Emergency protocols can temp reassign any agent any role
    • Cross-training → cover adjacent roles

→ Roles where agents progress simple→complex, distribution reflects needs+phase.

If err: rigid silos → ↑cross-training + rotation freq. Newcomers struggle progress → mentoring insufficient — pair w/ specialist for first N tasks. Too many in one role → triggers miscalibrated — adjust by colony-wide demand.

Step 3: Restructure Coord for Scale

Adapt mechanisms from coordinate-swarm for size.

  1. Replace direct comms → layered signaling:
    • Founding: everyone→everyone (N×N)
    • Growth: cluster squads of 5-8; direct in squad, signal between
    • Maturity: squads → departments; intra-squad direct, inter-squad signal, inter-dept broadcast
  2. Coord layers:
    • Local: in squad, direct signal exchange (stigmergy)
    • Regional: between squads same dept, aggregated signals
    • Colony: between depts, broadcast only for colony-wide decisions
  3. Inter-layer interfaces:
    • Each squad has 1 designated communicator who aggregates+relays
    • Communicators filter noise: not every local signal relayed up
    • Colony broadcasts rare → quorum, alarm escalation, major state changes
  4. Comms overhead budget:
    • Target: each agent <20% capacity on coord
    • Measure actual; exceed → add layer or split oversized squad

→ Layered coord, comms overhead grows logarithmic (not linear) w/ size. Local fast direct; colony-wide slower but functional.

If err: layers create info bottlenecks (communicators overloaded) → add redundant communicators or ↓relay freq. Layers create isolation (squads don't know others) → ↑inter-layer signal freq or cross-squad liaison roles.

Step 4: Execute Budding (Fission)

Split colony → semi-autonomous sub-colonies when exceeds single-coord capacity.

  1. Fission triggers:
    • 100 agents (or coord layer count >3)

    • Comms overhead >30% capacity despite layering
    • Decision latency exceeds time-sensitive thresholds
    • Subgroups have distinct identities + can operate independent
  2. Plan fission:
    • Identify natural split lines (existing clusters, domain bounds, geo)
    • Each daughter has viable role distribution (can't split all specialists into one)
    • Each must have: ≥1 coordinator, sufficient workers, access to shared resources
    • Define inter-colony interface: what shared, what independent
  3. Execute split:
    • Announce plan + timeline (consensus required — see build-consensus)
    • Transfer agents → daughters by existing cluster membership
    • Establish inter-colony channels (lightweight, async)
    • Each daughter bootstraps own local coord (inheriting from parent)
  4. Post-fission stabilization:
    • Monitor each for viability (sustains itself?)
    • Inter-colony coord minimal (quarterly sync, not daily)
    • Failed daughter → reabsorb into nearest viable

→ ≥2 viable daughters semi-autonomous w/ own coord, connected by lightweight interfaces.

If err: daughters too small → fission premature; remerge + retry larger. Inter-colony coord as heavy as pre-fission → split lines wrong, too interdependent. Re-draw on natural independence.

Step 5: Monitor Limits + Adapt

Continuous assess: structure matches size+needs?

  1. Scaling health metrics:
    • Coord overhead ratio: time coord/time produce
    • Decision throughput: decisions/time (↑ or steady w/ growth)
    • Agent satisfaction: engagement, retention, purpose (drops on fail)
    • Err rate: coord fails/time (not linear w/ growth)
  2. Limit indicators:
    • Overhead ratio >25% → more automation or layer
    • Throughput declining → governance needs revision
    • Turnover spiking → cultural|structural issues
    • Err rate accelerating → coord failing
  3. Trigger adapt:
    • Phase transition → apply Step 1 strategy
    • Limit reached → escalate (role diff → coord restructure → fission)
    • External change (market, tech) → may need transformation (see adapt-architecture)

→ Colony monitors own health + proactively adapts before stress = failure.

If err: no metrics → lacks observability — build measurement before more structure. Metrics show problems but can't adapt → resistance cultural not technical — address human factors (fear, ownership, trust) before restructure.

Check

  • Phase ID'd w/ specific stress signals
  • Role diff defined w/ progressive specialization
  • Coord layered for size
  • Comms overhead <20-25% capacity
  • Fission plan exists for >single-coord capacity
  • Health metrics tracked + thresholds trigger adapt
  • Daughter colonies (post-fission) viable distribution

Traps

  • Scale structure pre-needed: Premature layering = overhead w/o benefit. 10-team doesn't need dept coordinators. Stress signals guide.
  • Preserve founding culture at all costs: 5-agent ways break @ 50. Scaling needs evolution; nostalgia prevents adaptation.
  • Fission w/o independence: Sub-colonies still depend daily → worst of both — coord overhead + separation overhead.
  • Uniform role distribution: Not every sub-colony needs same ratios. Research → more scouts; production → more workers.
  • Ignore remerge: Sometimes fission fails; remerge best move. Treating fission irreversible prevents recovery.

  • coordinate-swarm — foundational patterns this skill scales
  • forage-resources — scales diff than production; role diff affects scout alloc
  • build-consensus — must adapt for larger groups
  • defend-colony — defense scales w/ colony
  • adapt-architecture — morphic skill for structural transformation
  • plan-capacity — capacity planning for growth
  • conduct-retrospective — identify stress before failure

GitHub リポジトリ

pjt222/agent-almanac
パス: i18n/caveman-ultra/skills/scale-colony
0
agentsagentskillsai-assisted-developmentclaude-codeskillsteams

関連スキル

llamaguard

その他

LlamaGuardは、暴力やヘイトスピーチなど6つの安全性カテゴリーにおいて、LLMの入力と出力をモデレートするMetaの70-80億パラメータモデルです。94〜95%の精度を提供し、vLLM、Hugging Face、Amazon SageMakerを使用してデプロイ可能です。このスキルを使用して、AIアプリケーションにコンテンツフィルタリングと安全策を簡単に統合できます。

スキルを見る

cost-optimization

その他

このClaudeスキルは、リソースの適正サイジング、タグ付け戦略、支出分析を通じて、開発者がクラウドコストを最適化することを支援します。AWS、Azure、GCPにわたるクラウド支出の削減とコストガバナンスの実施のためのフレームワークを提供します。インフラコストの分析、リソースの適正サイジング、または予算制約への対応が必要な際にご利用ください。

スキルを見る

quantizing-models-bitsandbytes

その他

このスキルは、bitsandbytesを使用してLLMを8ビットまたは4ビット精度に量子化し、精度の低下を最小限に抑えつつ50〜75%のメモリ削減を実現します。限られたGPUメモリでより大規模なモデルを実行したり、推論を高速化するのに理想的で、INT8、NF4、FP4などのフォーマットをサポートしています。HuggingFace Transformersと統合され、QLoRAトレーニングや8ビットオプティマイザーを可能にします。

スキルを見る

dispatching-parallel-agents

その他

このClaudeスキルは、複数のエージェントを配備し、3つ以上の独立した問題を並行して調査・修正します。共有状態や依存関係がなく解決可能な、無関係な障害が発生するシナリオ向けに設計されています。中核となる機能は並列問題解決であり、効率を最大化するために独立した問題領域ごとに1つのエージェントを割り当てます。

スキルを見る