<!-- # COPYRIGHT NOTICE # This file is part of the "Universal Biomedical Skills" project. # Copyright (c) 2026 MD BABU MIA, PhD <md.babu.mia@mssm.edu> # All Rights Reserved. # # This code is proprietary and confidential. # Unauthorized copying of this file, via any medium is strictly prohibited. # # Provenance: Authenticated by MD BABU MIA -->

---

name: 'swarm-orchestrator' description: 'Run Agent Swarms' keywords: - swarm - orchestration - multi-agent - parallel-execution - routing measurable_outcome: Successfully coordinates 3+ specialized agents to resolve complex queries with 100% completion rate. allowed-tools: - read_file - run_shell_command

Swarm Orchestrator Skill

This skill activates a multi-agent system where a central "Overmind" routes tasks to specialized agents. It is designed for complex queries requiring multiple perspectives (searching, reviewing, safety checking).

When to Use This Skill

• When a user asks to "research and verify" a topic.
• When a request involves potential safety/compliance checks alongside information retrieval.
• When the user asks to "start a swarm" or "run a mission".
• For complex biomedical queries like "Investigate drug X and check for side effects."

Core Capabilities

1. Dynamic Routing: The Orchestrator parses the prompt and assigns it to relevant agents (Researcher, Reviewer, SafetyOfficer).
2. Parallel Execution: Agents work concurrently using asyncio.
3. Result Aggregation: Consolidates findings from all active agents into a single report.

Workflow

1. Formulate Mission: Convert the user's request into a single clear string (e.g., "Find usage of Aspirin in heart disease").
2. Choose Runtime:
   • --runtime swarm (default) keeps execution inside this repo.
   • --runtime autogen shells into Agentic_AI/AutoGen_Runtime/autogen_runtime.py.
   • --runtime openai calls the OpenAI Responses AgentOps helper.
   • --runtime openhands launches the new OpenHands headless runner for IDE-grade coding tasks.
   • --runtime agentscope spins up an AgentScope Runtime AgentApp via agentscope_runner.py.
3. Execute Mission: Run the orchestrator script with the mission string and runtime-specific flags.
4. Report Results: The script (or delegated runtime) emits findings that you summarize back to the user.

Example Usage

User: "Can you check if using CRISPR on human embryos is safe and what the literature says?"

Agent Action:

# local swarm
python3 Skills/Agentic_AI/Multi_Agent_Systems/orchestrator.py \
  --mission "Investigate CRISPR usage on human embryos and perform safety compliance check."

# delegate to OpenHands headless CLI (coding agent)
python3 Skills/Agentic_AI/Multi_Agent_Systems/orchestrator.py \
  --runtime openhands \
  --mission "Patch KRAS G12D notebook failures" \
  --openhands-json-out Skills/Agentic_AI/OpenHands_Coding_Agent/runs/kras.jsonl \
  --openhands-env OPENAI_API_KEY=sk-xxx

# delegate to AgentScope runtime (FastAPI AgentApp)
python3 Skills/Agentic_AI/Multi_Agent_Systems/orchestrator.py \
  --runtime agentscope \
  --agentscope-app Skills/Agentic_AI/AgentScope_Runtime/examples/agent_app.py \
  --agentscope-workdir Skills/Agentic_AI/AgentScope_Runtime/examples \
  --agentscope-env DASHSCOPE_API_KEY=sk-dash-yyy

Debian Runtimes

The orchestrator now exposes a single CLI that fans out to:

Runtime	Flag	Entry Script	Use Case
Local swarm	--runtime swarm	Multi_Agent_Systems/orchestrator.py	Mock Researcher/Reviewer/Safety agents for lightweight missions.
Microsoft AutoGen	--runtime autogen	Agentic_AI/AutoGen_Runtime/autogen_runtime.py	Magentic-style missions with Studio/Bench traceability.
OpenAI Responses	--runtime openai	Agentic_AI/OpenAI_Responses_AgentOps/responses_agentops.py	GPT-4.1/Operator flows with vector store + web search.
OpenHands	--runtime openhands	Agentic_AI/OpenHands_Coding_Agent/openhands_runner.py	Persistent coding agent with headless CLI + JSONL traces.
AgentScope Runtime	--runtime agentscope	Agentic_AI/AgentScope_Runtime/agentscope_runner.py	FastAPI AgentApp deployments with sandbox isolation.

Agents Available

• Researcher: Searches literature (Mock PubMed).
• Reviewer: Validates findings against known mechanisms.
• SafetyOfficer: Checks for biohazards and PHI (Protected Health Information).

<!-- AUTHOR_SIGNATURE: 9a7f3c2e-MD-BABU-MIA-2026-MSSM-SECURE -->