Agent Swarm

What This Skill Does

Agent Swarm is a sophisticated traffic controller for the OpenClaw ecosystem, designed to intelligently route tasks to the most appropriate AI models via OpenRouter. Instead of relying on a single, static model for every request, this skill analyzes incoming tasks, selects an optimized model from the OpenRouter catalog, and automatically orchestrates the creation of sub-agent sessions using sessions_spawn. It acts as an intelligent middleware layer that balances complexity against compute, ensuring that specialized tasks (like coding or creative writing) are handled by models best suited for those specific workflows.

Installation

To integrate Agent Swarm into your environment, use the OpenClaw command-line interface: clawhub install openclaw/skills/skills/runeweaverstudios/agent-swarm. Before running, ensure your OpenClaw platform settings contain a valid OpenRouter API key. While the OPENCLAW_HOME environment variable is optional, the skill requires read access to your local openclaw.json file to identify tools.exec.host and tools.exec.node parameters. Please review your security posture before granting this read permission.

Use Cases

Agent Swarm is ideal for users with multi-faceted workflows who need to delegate tasks. It is particularly powerful for:

• Managing complex multi-step projects where specific segments require distinct specialized AI architectures.
• Automating the distribution of workload across multiple sub-agents to parallelize tasks like debugging and documentation generation.
• Dynamically choosing cost-effective or high-performance models based on the complexity of the current prompt.

Example Prompts

1. "Agent Swarm, please analyze my current repository and suggest a model for refactoring the authentication module, then spawn a session to implement it."
2. "I have a complex request involving both bug fixing and writing a technical summary; please use the multi-task router to handle these separately."
3. "List all available OpenRouter models supported by the current agent configuration and evaluate the best fit for generating a poem about coding."

Tips & Limitations

• Model Prefixing: Always ensure your selected models use the openrouter/ prefix, or the delegation process will fail.
• Security: When integrating this into your own scripts, never construct shell strings from raw user input. Always use list-based arguments with subprocess.run to prevent shell injection vulnerabilities.
• Read Access: The skill is strictly scoped to reading specific execution hosts/nodes from openclaw.json. It will never access your secret gateway keys or API tokens located in other config files.