OpenClaw Agent Orchestrator v2.0

Multi-agent orchestration for AI agent teams. 5 tools. Parallel execution. Automatic result aggregation. Zero infrastructure.

What's New in v2.0

• FleetCommand Protocol -- Intelligent task routing that sends tasks to the agent most qualified to handle them based on role, current load, and historical performance.
• Parallel Execution -- Spawn up to 8 agents simultaneously (Pro). Dispatch to all of them in parallel. Aggregate when done.
• AgentHealth Monitor -- Each spawned agent is health-checked every 30 seconds. Crashed agents are auto-restarted with their last known state.
• ResultMerge Engine -- Aggregate results from 2–8 agents into a coherent synthesized response. R1 resolves conflicts between agents.

Quick Start

{
  "openclaw-agent-orchestrator": {
    "type": "streamable-http",
    "url": "https://openclaw-agent-orchestrator-mcp.yagami8095.workers.dev/mcp"
  }
}

Add to Claude Desktop, Cursor, Windsurf, VS Code, or any MCP-compatible client. Zero install. Works immediately.

Detection Triggers

This skill activates when a user says any of these:

• "spawn a sub-agent", "create a worker agent", "delegate to agent", "launch agent for this"
• "list all agents", "what agents are running", "show active agents", "agent fleet status"
• "dispatch this task", "send to best agent", "route this to the right worker"
• "what is agent X doing", "agent status", "is it done", "agent progress", "check on agent"
• "aggregate results", "collect outputs", "merge agent answers", "synthesize responses"
• "parallel agents", "run these tasks in parallel", "concurrent agent execution"

Named Protocols

FleetCommand Protocol

The intelligent routing engine for multi-agent task dispatch:

Routing algorithm:

1. Role matching -- Task is tagged with required capability (web, code, analysis, writing)
2. Load check -- Agents at >80% capacity are deprioritized
3. Performance score -- Agents are ranked by task completion rate and avg latency (last 24h)
4. Selection -- Highest-score available agent with matching role receives the task
5. Fallback -- If no role-matched agent exists, task goes to general-purpose agent or spawns new one

Agent roles:

AgentHealth Monitor

Continuous health tracking for all spawned agents:

ResultMerge Engine

How aggregate_results synthesizes multi-agent outputs:

1. Collect -- Gather outputs from all specified agent IDs
2. Conflict detect -- R1 identifies contradictions, disagreements, or gaps between agents
3. Resolve -- R1 reasons through conflicts using confidence scoring and source weight
4. Synthesize -- Produces a single coherent response with attribution footnotes
5. Confidence flag -- If agents disagree >40% on a point, flags it as contested in output

Tools (5)

spawn_agent (Pro) -- Launch Agent via FleetCommand Protocol

Create a new agent with a specified role, model, and initial task. Returns agent ID for subsequent dispatch and status calls.

Wrong / Right:

WRONG: Running 5 tasks sequentially in one agent, waiting 10min for results
RIGHT: spawn_agent({ role: "analyst", model: "deepinfra/deepseek-ai/DeepSeek-R1-0528",
                     task: "Analyze Q1 revenue trends in our Stripe data" })
       spawn_agent({ role: "writer", task: "Draft the investor update email" })
       spawn_agent({ role: "scraper", task: "Pull competitor pricing from 5 sites" })
       -> All 3 run in parallel. Done in 3min instead of 10min.

WRONG: Spawning 20 agents for a simple 2-step task
RIGHT: Spawn agents only when parallelism provides actual speedup.
       Single sequential tasks are faster in one agent session.

list_agents -- Fleet Status

List all active agents with their current status, assigned role, current task, uptime, and resource usage.

Wrong / Right:

WRONG: Not knowing which agents are running or what they're doing
RIGHT: list_agents()
       -> [{ id: "agent-7f3a", role: "analyst", status: "running",
              task: "Analyzing Stripe data", started: "14min ago", cpu_pct: 42 },
           { id: "agent-2b8c", role: "writer", status: "idle", ... }]
       -> Full fleet visibility in one call

WRONG: Assuming agents are running when they may have crashed
RIGHT: Check list_agents before dispatching critical tasks.
       AgentHealth Monitor auto-restarts crashed agents, but there can be a 30–60s gap.

dispatch_task -- Task Routing via FleetCommand Protocol

Send a task to a specific agent ID, or let FleetCommand route it to the best available agent.

Parameters: task (string), agent_id (optional — omit for auto-routing), role_hint (optional), priority (low/normal/high)

get_agent_status -- Agent Detail

Get detailed status for a specific agent: current task, progress percentage, errors, estimated completion time, and last 5 log lines.

Returns: agent_id, role, status, current_task, progress_pct, errors (last 3), eta_sec, uptime_sec, last_heartbeat

aggregate_results -- Multi-Agent Result Synthesis via ResultMerge Engine

Collect and synthesize outputs from multiple agents. Pass a list of agent IDs and optionally a synthesis instruction.

Parameters: agent_ids (array), synthesis_instruction (optional), wait_for_completion (bool, default true), timeout_sec (default 120)

Security & Privacy

• Agent isolation -- Each spawned agent runs in its own isolated execution context. Agents cannot read each other's memory or task history.
• No credential sharing -- Agents spawned via this server do not inherit API keys or tokens from your session. You must explicitly pass credentials to each agent.
• Task content not logged -- Task content is processed in-memory and not stored server-side. Only metadata (agent ID, role, status) is persisted.
• Fleet visibility scoped -- list_agents only shows agents spawned under your Pro key. You cannot see other users' agents.
• Auto-cleanup -- Completed agents are terminated and purged after 24 hours. Crashed agents after 1 hour.
• HTTPS only -- All connections are TLS 1.3. HTTP is rejected.

Pricing

Get Pro Key: https://buy.stripe.com/4gw5na5U19SP9TW288

The OpenClaw Intelligence Stack

All 9 servers share one Pro key. $9/mo = 49 tools.