# New Agent

> Scaffold a new AG2 ConversableAgent with tool functions and LLM config. Use when creating a standalone agent that needs tools, a system prompt, and proper structure.

- Skill: `ag2ai/new-agent` (Agent Skill)
- Install (CLI): `npx skillmds@latest add ag2ai/new-agent`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ag2ai/new-agent/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: ag2ai (https://skillmd.com/u/ag2ai)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/ag2ai/new-agent

---


You are scaffolding a new AG2 (AutoGen) agent. Follow the AG2 framework patterns exactly.

## Instructions

1. Ask the user for:
   - Agent name and purpose
   - What tools/capabilities it needs
   - Which LLM model to use (default: gpt-4o-mini)
   - Whether it needs external API access

2. Create the agent following this exact structure:

### File Structure
```
agents/<agent-name>/
  <agent_name>.py    # Agent definition + tools
  README.md          # Capabilities documentation
```

### Agent Code Pattern

```python
import json
from autogen import ConversableAgent, LLMConfig
from autogen.tools import tool

# --- Tool Functions ---
# Each tool returns a JSON string with {"success": bool, "data": ..., "error": ...}

@tool()
def tool_name(param1: str, param2: int = 10) -> str:
    """Clear description of what this tool does.

    Args:
        param1: Description of param1
        param2: Description of param2 (default: 10)
    """
    try:
        # Implementation
        result = {"key": "value"}
        return json.dumps({"success": True, "data": result})
    except Exception as e:
        return json.dumps({"success": False, "error": str(e)})


# --- Agent Definition ---
agent = ConversableAgent(
    name="agent_name",
    description="One-line description for orchestrator routing",
    system_message="""You are a [role description].

Your capabilities:
- Capability 1
- Capability 2

Guidelines:
- Always use the appropriate tool for the task
- Return structured responses
- Handle errors gracefully and explain what went wrong
""",
    llm_config=LLMConfig({"api_type": "anthropic", "model": "claude-sonnet-4-6"}),
    functions=[tool_name],
)
```

### Key Rules

- Tool functions MUST have docstrings (used for LLM function calling schema)
- System messages should be specific about the agent's role and boundaries
- Agent `description` is used by orchestrators to route tasks -- keep it concise
- Use `@tool()` decorator from `autogen.tools`
- Group related tools in the same file
- Never use bare `except:` -- always catch specific exceptions or `Exception`

3. After scaffolding, verify the code is syntactically valid and all imports exist.