agent-creator_skill

Overview

This skill guides creation of Multi-Agent System (MAS) implementations for the AWorld framework, supporting both centralized (TeamSwarm) and decentralized (Swarm) coordination patterns. It automates the agent package layout and generates configuration, prompt, and swarm initialization files under ./agents/, ready for aworld_cli discovery and smoke tests. Use it to rapidly scaffold agent teams that rely on filesystem-server MCP tools and standard Agent class usage.

How this skill works

The skill inspects task requirements and recommends a MAS type (TeamSwarm or Swarm). It uses filesystem-server tools (write_file, edit_file, read_file) to create the directory structure, __init__.py, config.py, prompt.py, and swarm.py files under ./agents/{team_name}/. Swarm initialization files instantiate aworld.agents.llm_agent.Agent instances directly and register teams with the @agent decorator so aworld_cli can discover and run them. It enforces critical Python indentation rules and provides templates for common workflows.

When to use it

• You need a team with a central orchestrator that dynamically routes work to specialists (TeamSwarm).
• You have a fixed, sequential workflow where agents run in a defined order (Swarm).
• You want a quick, repeatable scaffold for agents that aworld_cli can discover and run.
• You need to define MCP servers and filesystem-server access per agent.
• You want to avoid custom agent classes and use the provided Agent class directly.

Best practices

• Analyze task complexity first: choose Swarm for simple pipelines, TeamSwarm for dynamic coordination.
• Keep top-level Python code at column 0 in generated files; always verify with read_file.
• Provide a concise system prompt in prompt.py for each role; config.py can be minimal or omitted.
• Instantiate Agent directly in swarm.py and register the team with @agent for automatic discovery.
• Smoke-test new agents with aworld-cli list and aworld-cli --task "Hello" --agent="YourTeamName".

Example use cases

• Research team with an orchestrator that delegates web browsing and data summarization tasks.
• Document generation pipeline: analysis → outline → content → HTML export using a Swarm.
• Single-file Markdown agent for quick document assistants loaded automatically by aworld_cli.
• Rapid prototyping of agent teams that require filesystem access via filesystem-server MCP.
• Upgrading a simple Swarm to a TeamSwarm when routing decisions become necessary.