name: orchestrator description: Multi-Agent Orchestration skill for coordinating AI agents across channels, managing sessions, distributing tasks, and leveraging distributed compute resources.
Orchestrator Skill
Purpose
Enable AI agents to act as Orchestrators - coordinating multiple AI agents across federated communication channels, managing sessions, distributing tasks, and aggregating results.
When to Use This Skill
Activate when you need to:
- Coordinate multiple AI agents for collaborative work
- Distribute tasks across parallel compute resources
- Manage multi-agent sessions with ID assignment
- Aggregate results from multiple sources
- Leverage free/low-cost AI compute (Gemini tabs, Jules CLI, etc.)
Quick Start
1. Start the Relay Server
pnpm relay:start
2. Launch Persistent Orchestrator
node orchestrator-persistent.js Green # Join Green channel
3. Session Control Commands
node session-control.js status # Check agent status
node session-control.js pause # Pause all agents
node session-control.js resume # Resume operations
node session-control.js end # End session
Protocol: DACC-v1
Distributed Agent Communication & Coordination Protocol v1
Core Principles
-
Mandatory Signing - All agents MUST sign messages with
[AGENT-XX] - ID Assignment - Orchestrator assigns unique IDs to agents
- Structured Communication - Use formatted message templates
- Session Logging - All conversations logged to markdown files
Message Signing Format
[AGENT-XX] Your message content here
Acceptable alternatives:
-
ID#: AGENT-XX message -
[GEMINI-3F-01] message(custom IDs)
Compute Resources
Available Sources (Prioritized by Cost)
| Source | Cost | Parallelism | Best For |
|---|---|---|---|
| Gemini Web Tabs | FREE | Multiple tabs | Analysis, reasoning |
| Jules CLI | FREE | Up to 4 parallel | Async coding tasks |
| HuggingChat | FREE | 1 | Open source models |
| Perplexity | FREE | 1 | Search + reasoning |
| Claude Free | FREE | Limited | Complex reasoning |
| ChatGPT Free | FREE | Limited | General tasks |
Jules CLI Integration
# Check Jules status
jules version
jules remote list --session
# Submit task
jules new "Implement feature X in packages/core"
# Parallel execution
jules new --parallel 4 "Add tests to all services"
# Pull completed work
jules remote pull --session 123456 --apply
Chrome Extension (Gemini Tabs)
- Open multiple Gemini tabs
- Press Ctrl+Shift+F to open Fuse Connect
- Join the same channel in each tab
- Each tab becomes an independent agent
Session Management
Starting a Session
-
Start Relay:
pnpm relay:start -
Launch Orchestrator:
node orchestrator-persistent.js [channel] - Open Agent Tabs: Gemini tabs with Fuse extension
- Join Channel: All agents join same channel
Session Commands
| Command | Script | Purpose |
|---|---|---|
| Pause | node session-control.js pause |
Request all agents to pause |
| Resume | node session-control.js resume |
Resume paused session |
| Status | node session-control.js status |
Request status from all agents |
| End | node session-control.js end |
End the current session |
Session Logs
All sessions are logged to:
.agent/session-logs/YYYY-MM-DDTHH-MM-SS_Channel_persistent.md
Log contents:
- Session metadata (ID, times, channel, protocol)
- Participant table (assigned IDs, message counts, violations)
- Full conversation log with timestamps
- Signing violation tracking
Task Distribution
Task Template
═══════════════════════════════════════════════════════════════
🔧 TASK: [Task Name]
═══════════════════════════════════════════════════════════════
Task ID: [unique-id]
Priority: [high|medium|low]
Assigned To: [AGENT-XX or "all"]
📋 DESCRIPTION
[Clear description of what needs to be done]
📌 DELIVERABLES
1. [Deliverable 1]
2. [Deliverable 2]
⏱️ DEADLINE: [Time constraint if any]
🔔 Sign responses with [AGENT-XX]!
═══════════════════════════════════════════════════════════════
Parallel Task Distribution
ORCHESTRATOR assigns:
├── AGENT-01: Analyze packages/a2a-core
├── AGENT-02: Analyze packages/agent
├── JULES-01: Implement security fixes (async)
└── JULES-02: Add unit tests (async)
Scripts Reference
| Script | Purpose |
|---|---|
orchestrator-persistent.js |
Persistent orchestrator with auto-reconnect |
orchestrator-green-channel.js |
Single-session orchestrator |
session-control.js |
Pause/resume/end session commands |
send-improvement-task.js |
Send improvement tasks |
scripts/list_channels.js |
List available channels |
Self-Improvement Cycle
Daily Protocol
- Review - Check overnight agent conversations in session logs
- Identify - Find improvement candidates from agent analyses
- Prioritize - Rank by impact and effort
- Distribute - Assign tasks to available compute
- Implement - Agents execute tasks
- Verify - Check results and integrate
- Document - Update documentation and logs
Continuous Improvement File
Track improvement status in:
.agent/SELF_IMPROVEMENT_CYCLE.md
Integration with Other Skills
agent-discovery-protocol
Protocol specification for agent discovery and ID assignment.
jules-delegation
Delegate async coding tasks to Jules CLI.
chrome-devtools
Browser automation for testing and verification.
Troubleshooting
Agents Not Responding
- Check relay is running:
pnpm relay:start - Verify agents are on same channel
- Check orchestrator logs for errors
Signing Violations
- Orchestrator sends reminders automatically
- Check session log for violation count
- Send manual reminder:
node session-control.js status
Connection Drops
- Persistent orchestrator auto-reconnects in 5 seconds
- Check relay server health:
curl localhost:3001/health - Agents should auto-reconnect on tab refresh
Issue Resolution Protocol
Orchestrators MUST use available tools to resolve issues autonomously.
When Issues Arise
- Identify the issue type
- Select appropriate tool(s)
- Execute the fix
- Verify the resolution
- Document what was done
Tool Selection Matrix
| Issue Type | Tools to Use | Action |
|---|---|---|
| Build failure | run_command, grep_search, view_file |
Find error, view code, fix |
| Test failure | run_command, view_file, replace_file_content |
Run tests, view failures, fix |
| Agent not responding | mcp_chrome-devtools_*, run_command |
Check browser, restart relay |
| Connection error | run_command, read_terminal |
Check relay status, restart |
| Missing file | write_to_file, run_command |
Create file, verify |
| Code error | view_file, replace_file_content |
View, fix, verify |
| Documentation gap | write_to_file, replace_file_content |
Update docs |
Auto-Resolution Prompts
When encountering an error, use this mental framework:
I encountered: [ERROR DESCRIPTION]
Available tools that could help:
1. run_command - Execute shell commands to diagnose/fix
2. view_file - Inspect source code for issues
3. grep_search - Find related code or error patterns
4. replace_file_content - Fix code issues
5. write_to_file - Create missing files
6. browser_subagent - Interact with browser for testing
7. mcp_chrome-devtools_* - Direct browser control
I will:
1. Use [TOOL] to [ACTION]
2. Verify with [VERIFICATION METHOD]
3. Document the fix in session log
Common Issue Resolutions
Relay Not Running
# Check status
curl localhost:3001/health
# If down, restart
pnpm relay:start
Agent Not Signing Messages
Send reminder via session-control.js status
or broadcast direct reminder to channel
Build Errors
1. run_command: pnpm build 2>&1 | head -50
2. grep_search: Find the error file
3. view_file: Inspect the error location
4. replace_file_content: Fix the issue
5. run_command: Verify build passes
Jules Task Failed
# Check session status
jules remote list --session
# Get details
jules remote pull --session [ID]
# Retry with improved prompt
jules new "Fixed version of: [original task]"
Escalation
If tools cannot resolve the issue:
- Document the issue fully in session log
- Create a clear summary for human review
- Continue with other tasks that can proceed
Version History
-
v1.0.0 (2026-01-16)
- Initial orchestrator skill
- Persistent mode with auto-reconnect
- Session control commands (pause/resume/end)
- Compute resource documentation
- DACC-v1 protocol integration
chat Comments (0)
Sign in to join the discussion and leave a comment.
Skill Details
Related Skills
Build your own?
Join 12,000+ developers contributing to the Claude ecosystem.
No comments yet. Be the first to share your thoughts!