AgentSkillsCN

swarm

协调多代理编排复杂任务。启动并行和串行代理，管理依赖关系，汇总结果，编排复杂工作流。用于需要多个专业视角或并行处理的任务。

SKILL.md
--- frontmatter
name: swarm
description: Coordinate multi-agent orchestration for complex tasks. Launch parallel and sequential agents, manage dependencies, aggregate results, and orchestrate sophisticated workflows. Use for tasks requiring multiple specialized perspectives or parallel processing.
license: MIT

ELF Swarm Coordination Command

Orchestrate multi-agent workflows for complex tasks requiring parallel processing or multiple specialized perspectives.

Purpose

The /swarm command enables:

  • •Parallel processing - Multiple agents working simultaneously
  • •Specialized perspectives - Researcher, Architect, Creative, Skeptic agents
  • •Dependency management - Sequential processing when needed
  • •Result aggregation - Combine outputs from multiple agents
  • •Sophisticated workflows - Complex orchestration patterns

Usage Examples

code
/swarm analyze my architecture from 4 perspectives
/swarm run parallel searches on [topics]
/swarm investigate this failure through agent lenses
/swarm parallelize this migration task

Key Agent Perspectives

Researcher

  • •Asks: "What does the evidence say?"
  • •Strength: Finds authoritative knowledge
  • •Use: For data-driven decisions

Architect

  • •Asks: "How does this scale?"
  • •Strength: Systems thinking
  • •Use: For structural decisions

Creative

  • •Asks: "What if we tried something different?"
  • •Strength: Novel solutions
  • •Use: When stuck on problems

Skeptic

  • •Asks: "What could go wrong?"
  • •Strength: Finds edge cases
  • •Use: For validation

How Swarm Orchestration Works

When you invoke /swarm:

  1. •Parse your request - Understand task and constraints
  2. •Plan execution - Determine parallel vs sequential
  3. •Launch agents - Spawn subagents in background
  4. •Manage dependencies - Block only when needed
  5. •Aggregate results - Combine perspectives
  6. •Synthesize insights - Extract unified understanding

Swarm Patterns

Parallel Analysis

Launch all 4 agents simultaneously on the same problem. Best for: Complex decisions, design reviews, failure analysis

Sequential Pipeline

Run agents in sequence where each builds on previous. Best for: Iterative refinement, progressive investigation

Expert Consultation

Launch specific agents for their expertise. Best for: Targeted investigation

Parallel + Synthesis

Run multiple agents in parallel, then synthesize results. Best for: Comprehensive analysis

Agent Coordination Rules

  • •Always run in background - run_in_background=True
  • •Block only when needed - Use TaskOutput to wait for results
  • •Specify agent type - Researcher, Architect, Creative, Skeptic
  • •Use models efficiently - Haiku for small tasks, Sonnet/Opus for complex
  • •Aggregate thoughtfully - Synthesize perspectives, don't just list them

Integration with ELF

Swarm results can feed back into the building:

  • •Document learnings - Record what agents discovered
  • •Update heuristics - If swarm validates/challenges existing knowledge
  • •Propose rules - If discovery is universal enough
  • •Escalate decisions - If swarm surfaces ambiguity

Example Workflow

code
1. User: "/swarm analyze my architecture from 4 perspectives"
2. System: Launches 4 agents in parallel
   - Researcher: Evidence-based evaluation
   - Architect: Structural analysis
   - Creative: Alternative approaches
   - Skeptic: Risk identification
3. System: Aggregates results into synthesis
4. User: Gets comprehensive perspective
5. Building: Results documented if significant

When to Use Swarm

  • •Complex decisions - Need multiple viewpoints
  • •Ambitious goals - Parallel processing helps
  • •Risk management - Skeptic finds what you missed
  • •Stuck problems - Creative breaks conventional thinking
  • •Learning opportunities - Results feed building's knowledge