AgentSkillsCN

Agent Coordination

智能体协调

SKILL.md

Agent Coordination Skill

Multi-agent coordination for Claude Code sessions using shared state and structured communication.

Overview

This skill enables multiple Claude Code agents (subagents) to work together on complex tasks by sharing findings, coordinating work, and avoiding duplication of effort.

When to Use

Use agent coordination when:

  • •A task has multiple independent parts that can be parallelized
  • •You need different perspectives (investigation, design, critique)
  • •The problem benefits from divide-and-conquer
  • •You're spawning subagents with the Task tool

Core Concepts

The Blackboard

Agents coordinate through a shared "blackboard" - a file-based state store at .coordination/blackboard.json in the project root. The blackboard tracks:

  • •Agent Registry - Who is working on what
  • •Findings - Discoveries, warnings, blockers agents want to share
  • •Messages - Direct agent-to-agent communication
  • •Task Queue - Pending work items
  • •Questions - Blockers needing resolution

Agent Personas

Four specialized personas for different thinking styles:

PersonaRoleWhen to Use
ResearcherDeep investigation, gathering evidence"We need to understand X"
ArchitectSystem design, big picture planning"How should we structure X?"
SkepticBreaking things, finding edge cases"Is this robust?"
CreativeNovel solutions, alternative approaches"We're stuck on X"

Personality definitions are in ~/.claude/emergent-learning/agents/*/personality.md.

Finding Types

When sharing discoveries, use these types:

  • •discovery - Found something interesting
  • •warning - Potential issue to be aware of
  • •decision - A choice that was made
  • •blocker - Something preventing progress
  • •fact - Verified information
  • •hypothesis - Unverified theory to test

Setup

1. Initialize Coordination Directory

Copy the templates to your project:

bash
mkdir -p .coordination
cp ~/.claude/skills/agent-coordination/TEMPLATES/* .coordination/

2. Include in Subagent Prompts

When spawning subagents, include coordination instructions:

code
You are working as part of a coordinated team.

Before starting:
1. Register yourself: python .coordination/blackboard.py register "<agent-id>" "<task>"
2. Check for existing findings: python .coordination/blackboard.py summary

During work:
- Add findings: python .coordination/blackboard.py finding "<agent-id>" "<type>" "<content>"
- Check for updates: python .coordination/blackboard.py delta "<agent-id>"

When done:
- Update status with final result

API Reference

Blackboard CLI

bash
# View current state
python .coordination/blackboard.py summary

# Register as working agent
python .coordination/blackboard.py register <agent-id> "<task>" [interest-tags...]

# Add a finding
python .coordination/blackboard.py finding <agent-id> <type> "<content>" [tags...]

# Search findings (keyword)
python .coordination/blackboard.py search "<query>"

# Check what's new since last check
python .coordination/blackboard.py delta <agent-id>

# Reset blackboard
python .coordination/blackboard.py reset

Python API

python
from blackboard import Blackboard

bb = Blackboard(project_root=".")

# Register
bb.register_agent("agent-1", "Investigate auth system", interests=["auth", "security"])

# Add finding
bb.add_finding("agent-1", "discovery", "Found unused JWT validation",
               files=["src/auth.py"], importance="high", tags=["jwt", "security"])

# Get findings relevant to interests
relevant = bb.get_findings_for_interests(["security"])

# Get new findings since last check
cursor = bb.get_agent_cursor("agent-1")
new_findings = bb.get_findings_since_cursor(cursor)
bb.update_agent_cursor("agent-1")

Integration with Basic Memory (Optional)

For semantic search across findings, you can optionally use Basic Memory MCP tools (requires Basic Memory MCP server):

python
# Write finding to Basic Memory for semantic search
mcp__basic-memory__write_note(
    title="Discovery: JWT validation unused",
    content="Found that jwt_validate() is never called...",
    folder="coordination",
    project="my-project"
)

# Semantic search across all findings
mcp__basic-memory__search_notes(
    query="authentication security issues",
    project="my-project"
)

# Build context from coordination folder
mcp__basic-memory__build_context(
    url="coordination/*",
    project="my-project"
)

The blackboard handles real-time coordination; Basic Memory provides persistent semantic search.

Best Practices

  1. •Register before working - Let other agents know what you're doing
  2. •Share findings early - Don't wait until done to share discoveries
  3. •Use appropriate finding types - Helps other agents prioritize
  4. •Tag findings - Makes filtering and searching easier
  5. •Check for updates - Periodically check what others have found
  6. •Mark blockers clearly - Use blocker type and critical importance

Limitations

  • •Pro/Max plan required - Free plan can't use Task tool for subagents
  • •File-based locking - Works for moderate concurrency, not high-throughput
  • •No built-in semantic search - Use Basic Memory MCP for that capability
  • •Manual cursor management - Agents must track their own position

Files

code
~/.claude/skills/agent-coordination/
  TEMPLATES/            # Template files for coordination
  SKILL.md              # This documentation
  hooks/
    hooks.json          # Hook definitions
    pre_task.py         # Pre-task context injection
    post_task.py        # Post-task finding recording
    session_end.py      # Session cleanup
  utils/
    blackboard.py       # Core coordination logic
    sqlite_bridge.py    # Database integration

.coordination/          # Per-project (in project root)
  blackboard.json       # Shared state file
  .blackboard.lock      # Lock file for concurrency

See Also

  • •Swarm Commands - /swarm slash command for orchestrated multi-agent work
  • •Conductor - Workflow orchestration for complex tasks
  • •Agent Personas - Personality definitions for each agent type