coordinate-swarm
О программе
Этот навык предоставляет шаблоны для децентрализованной координации в распределенных системах с использованием концепций роевого интеллекта, таких как стигмергия и кворумное зондирование. Он помогает разработчикам проектировать системы, в которых автономные агенты самоорганизуются посредством локальных правил и общих сигналов вместо централизованного управления. Используйте его при создании устойчивых событийно-ориентированных архитектур или для замены хрупкой оркестровки на эмерджентную координацию.
Быстрая установка
Claude Code
Рекомендуетсяnpx 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/coordinate-swarmСкопируйте и вставьте эту команду в Claude Code для установки этого навыка
Документация
Coordinate Swarm
Set up coordination across distributed agents. Use stigmergy (indirect talk through environment), local rules, quorum sensing. Coherent group behavior, no central boss.
When Use
- Designing distributed system, no single node should be bottleneck
- Organizing teams, workflows must self-coordinate, no constant boss oversight
- Building event-driven architecture, parts talk through shared state, not direct messages
- Scaling process works fine with 3 agents, breaks at 30
- Bootstrapping coordination for new swarm domain (see
forage-resources,build-consensus) - Replacing fragile central orchestration with resilient emergent coordination
Inputs
- Required: Description of agents (workers, services, team members) needing coordination
- Required: Collective goal or wanted emergent behavior
- Optional: Current coordination method and its failure modes
- Optional: Agent count (affects pattern choice — small swarm vs. big colony)
- Optional: Latency tolerance (real-time vs. eventual)
- Optional: Environment limits (shared state, bandwidth)
Steps
Step 1: Identify Coordination Problem Class
Classify challenge. Pick right patterns.
- Map current state: who are agents, what do they do alone, where does coordination break?
- Classify problem:
- Foraging — agents search for, exploit distributed resources (see
forage-resources) - Consensus — agents must agree on collective decision (see
build-consensus) - Construction — agents build, maintain shared structure step by step
- Defense — agents detect, respond to threats collectively (see
defend-colony) - Division of labor — agents must self-organize into specialized roles
- Foraging — agents search for, exploit distributed resources (see
- Identify failure mode of current coordination:
- Single point of failure (central controller)
- Communication bottleneck (too many direct messages)
- Coherence loss (agents drift apart, no feedback)
- Rigidity (cannot adapt to change)
Got: Clear classification of coordination problem type and specific failure mode to fix. Picks swarm patterns to use.
If fail: Problem does not fit single class? May be mixed. Break into sub-problems, address each with own pattern. Agents too heterogeneous for single model? Consider layered coordination — homogeneous clusters coordinated via inter-cluster stigmergy.
Step 2: Design Stigmergic Signals
Build indirect talk channels. Agents influence each other through them.
- Define shared environment (database, message queue, file system, physical space, shared board)
- Design signals agents drop into environment:
- Trail signals: markers pile up along good paths (like ant pheromones)
- Threshold signals: counters trigger behavior change when crossing threshold
- Inhibition signals: markers push agents away from exhausted areas
- Define signal properties:
- Decay rate: how fast signals fade (stops stale state dominating)
- Reinforcement: how good outcomes strengthen signals
- Visibility radius: how far signal spreads
- Map signals to agent behaviors:
- Agent senses signal X above threshold T → does action A
- Agent finishes action A well → drops signal Y
- No signal sensed → agent does default exploration
Signal Design Template:
┌──────────────┬───────────────────┬──────────────┬────────────────────┐
│ Signal Name │ Deposited When │ Decay Rate │ Agent Response │
├──────────────┼───────────────────┼──────────────┼────────────────────┤
│ success-trail│ Task completed OK │ 50% per hour │ Follow toward │
│ busy-marker │ Agent starts task │ On completion│ Avoid / pick other │
│ help-signal │ Agent stuck >5min │ 25% per hour │ Assist if nearby │
│ danger-flag │ Error detected │ 10% per hour │ Retreat & report │
└──────────────┴───────────────────┴──────────────┴────────────────────┘
Got: Signal table mapping environment markers to deposit conditions, decay rates, response behaviors. Signals simple, composable, each meaningful alone.
If fail: Signal design feels too complex? Drop to two signals: one positive (success trail), one negative (danger flag). Most coordination bootstraps with attract/repel. Add nuance only after basic system works.
Step 3: Define Local Interaction Rules
Write simple rules each agent follows. Only local info (own state + nearby signals).
- Define agent's perception radius (what can sense?)
- Write 3-7 local rules in priority order:
- Rule 1 (safety): Danger-flag sensed → move away
- Rule 2 (response): Help-signal sensed and idle → move toward
- Rule 3 (exploitation): Success-trail sensed → follow toward strongest signal
- Rule 4 (exploration): No signals sensed → move random, bias toward unexplored
- Rule 5 (deposit): After task done → drop success-trail at spot
- Each rule must be:
- Local: depends only on what agent sees
- Simple: one if-then statement
- Stateless (preferred): no need to remember past states
- Test rules in head: every agent follows these rules → wanted collective behavior emerges?
Got: Ranked rule set each agent runs alone. Applied across swarm → target collective behavior (foraging, construction, defense, etc.).
If fail: Mental sim does not produce wanted emergent behavior? Rules likely need feedback loop — agents must see consequences of collective actions. Add signal for collective state (e.g., "task completion rate") and rule adjusting behavior from it.
Step 4: Calibrate Quorum Sensing
Set thresholds triggering collective state changes when enough agents agree.
- Find decisions needing collective agreement (not just individual response):
- Switching from exploration to exploitation mode
- Committing to new work site or abandoning old one
- Escalating from normal to emergency response
- For each collective decision, define:
- Quorum threshold: count or percent of agents must signal agreement
- Sensing window: time period signals counted over
- Hysteresis: different thresholds for activation vs. deactivation (stops oscillation)
- Implement quorum as signal accumulation:
- Each agent favoring decision drops vote-signal
- Accumulated votes exceed quorum within sensing window → decision activates
- Votes drop below deactivation threshold → decision reverses
Got: Quorum thresholds let swarm make collective decisions, no leader. Hysteresis gap stops fast oscillation between states.
If fail: Swarm oscillates between states? Widen hysteresis gap (e.g., activate at 70%, deactivate at 30%). Swarm never reaches quorum? Lower threshold or grow sensing window. Decisions too slow? Shrink sensing window — but watch for premature consensus.
Step 5: Test and Tune Emergent Behavior
Validate local rules produce wanted collective behavior, then tune params.
- Run sim or pilot with small agent count (5-10)
- Observe:
- Swarm converges on intended behavior?
- How long convergence takes?
- What happens when conditions change mid-task?
- What happens when agents fail or get added?
- Tune params:
- Signal decay rate: too fast → no coordination memory; too slow → stale signals dominate
- Quorum threshold: too low → premature collective decisions; too high → paralysis
- Exploration-exploitation balance: too much exploration → wasteful; too much exploitation → local optima
- Stress test:
- Remove 30% of agents fast — swarm recovers?
- Double agent count — swarm still coordinates?
- Inject conflicting signals — swarm resolves or deadlocks?
Got: Tuned param set. Swarm self-organizes toward target behavior, recovers from shocks, scales clean.
If fail: Swarm fails stress tests? Signal design likely too tightly coupled. Simplify: fewer signals, higher decay rates (fresh info), agents have robust default behavior when no signals present. Swarm that does something reasonable with zero signals is more resilient than one depending on signal availability.
Checks
- Coordination problem classified into known pattern (foraging, consensus, construction, defense, division of labor)
- Stigmergic signal table defined: deposit conditions, decay rates, agent responses
- Local interaction rules simple, local, prioritized (3-7 rules)
- Quorum thresholds set with hysteresis to stop oscillation
- Small-scale test shows emergent behavior matching collective goal
- Stress test (agent removal, addition, signal disruption) shows graceful degradation
Pitfalls
- Over-engineering signals: Too many signal types upfront → confusion. Start with 2 signals (attract/repel), add only when proven needed
- Centralized thinking in disguise: "Local rule" needs agent to know global state → not local. Refactor until each rule depends only on what agent senses directly
- Ignoring decay: Signals never decaying → fossilized coordination state. Every signal needs half-life fitting task time scale
- Zero hysteresis: Quorum thresholds with no gap between activation and deactivation → fast state oscillation. Always set deactivation lower than activation
- Assuming homogeneity: Agents have different abilities → single rule set may fail. Consider role-differentiated rules (see
scale-colony)
See Also
forage-resources— applies swarm coordination to resource search and explore-exploit tradeoffsbuild-consensus— deep dive into distributed agreement, extends quorum sensing from this skilldefend-colony— collective defense patterns built on signal and rule framework herescale-colony— scaling when swarm outgrows initial coordination designadapt-architecture— morphic skill for transforming system architecture, pairs with swarm coordination triggering structural changedeploy-to-kubernetes— practical distributed system deployment where swarm patterns applyplan-capacity— capacity planning informed by swarm scaling dynamicscoordinate-reasoning— AI self-application variant; maps stigmergic signals to context management with info decay rates and local protocols
GitHub репозиторий
Похожие навыки
content-collections
МетаЭтот навык предоставляет проверенную в продакшене настройку для Content Collections — TypeScript-ориентированного инструмента, который преобразует файлы Markdown/MDX в типобезопасные коллекции данных с валидацией Zod. Используйте его при создании блогов, сайтов документации или контентных приложений на Vite + React для обеспечения типобезопасности и автоматической проверки содержимого. Он охватывает всё: от настройки плагина Vite и компиляции MDX до оптимизации развертывания и валидации схем.
polymarket
МетаЭтот навык позволяет разработчикам создавать приложения на платформе прогнозных рынков Polymarket, включая интеграцию с API для торговли и получения рыночных данных. Он также обеспечивает потоковую передачу данных в реальном времени через WebSocket для отслеживания текущих сделок и рыночной активности. Используйте его для реализации торговых стратегий или создания инструментов, обрабатывающих обновления рынка в реальном времени.
creating-opencode-plugins
МетаЭтот навык помогает разработчикам создавать плагины OpenCode, которые подключаются к более чем 25 типам событий, таким как команды, файлы и операции LSP. Он предоставляет структуру плагина, спецификации API событий и шаблоны реализации для модулей на JavaScript/TypeScript. Используйте его, когда вам нужно перехватывать, отслеживать или расширять жизненный цикл ассистента OpenCode AI с помощью пользовательской событийно-ориентированной логики.
sglang
МетаSGLang — это высокопроизводительный фреймворк для обслуживания больших языковых моделей (LLM), специализирующийся на быстрой структурированной генерации JSON, regex и рабочих процессов агентов с использованием кэширования префиксов RadixAttention. Он обеспечивает значительно более высокую скорость вывода, особенно для задач с повторяющимися префиксами, что делает его идеальным для сложных структурированных результатов и многократных диалогов. Выбирайте SGLang вместо альтернатив, таких как vLLM, когда вам требуется ограниченное декодирование или вы создаете приложения с интенсивным совместным использованием префиксов.
