Skip to content

Latest commit

 

History

History
49 lines (40 loc) · 1.83 KB

File metadata and controls

49 lines (40 loc) · 1.83 KB
name swarm-orchestrator
description Multi-agent orchestration using cooperative dynamics from arXiv:2602.16301. Spawns diverse agent swarms via task tool or independent copilot CLI subprocesses. Agents adapt through anonymous interaction history and converge on optimal solutions via mutual shaping. Use when tasks span 3+ files, need design decisions, or benefit from competing perspectives. Invoke with /swarm-orchestrator or "swarm this".
license MIT
metadata
author version paper mcp
Tim Schwarz
7.0
arXiv:2602.16301 — Wołczyk, Weis, Nasser et al. (2026)
open-swarm

Swarm Orchestrator — Launcher

When this skill is invoked, you have ONE job: launch the swarm-runner agent.

Do this IMMEDIATELY — no exploration, no planning, no codebase analysis:

task(
  agent_type="swarm-runner",
  description="Swarm: <brief summary>",
  prompt="<paste the user's COMPLETE request here verbatim — include every detail, objective, config, and constraint they provided>",
  mode="sync"
)

That's it. The swarm-runner agent handles everything:

  • Calls swarm_init on the Open Swarm MCP server
  • Gets workstream assignments, models, and phases from the server
  • Dispatches sub-agents via task() with server-assigned parameters
  • Manages merges, quality gates, and retries
  • Reports final results

Rules

  1. Do NOT call swarm_init yourself — the swarm-runner agent does that
  2. Do NOT launch multiple task() agents yourself — the swarm-runner manages sub-agents
  3. Do NOT create a plan or explore the codebase — pass the request through
  4. Do NOT modify or summarize the user's prompt — pass it VERBATIM
  5. If the user says "subprocess mode" or "independent agents", append that to the prompt

After swarm-runner completes

Report its output to the user. If it failed, you may retry once with a refined prompt.