create_subagent

Files (1)

• SKILL.md

SKILL.md (first lines)

# Create Subagent

Create a new Python subagent file in the **workspace** (temporary directory).

## IMPORTANT: Workflow After Creating

After `create_subagent`, use `run_subagent` to run it:
```
create_subagent → run_subagent (with query) → finish
```

**DO NOT** use `use_saved_subagent` for newly created subagents!
- `run_subagent`: For subagents you just created (in workspace)
- `use_saved_subagent`: ONLY for subagents listed by `list_saved_subagents`

## Usage

```xml
<action>create_subagent</action>
<params>
{
    "skill_name": "my_subagent_name",
    "filename": "subagent.py",
    "code": "your python code here",
    "skills": ["local_search", "open_page"]
}
</params>
```

## Parameters

| Parameter | Type | Default | Description |
|-----------|------|---------|-------------|
| skill_name | str | required | Name for this subagent (used when saving with `finish`) |
| filename | str | "subagent.py" | Name of the file to create in workspace |
| code | str | required | The Python code for the subagent |
| skills | list | ["local_search", "open_page"] | List of tool skill names this subagent will use. You MUST call `get_skill_description` for each tool skill BEFORE creating the subagent to understand how to use them. |

## Returns

```python
{
    "success": True,
    "path": "/path/to/workspace/subagent.py",
    "message": "Subagent created with skills: [...]"
}
```

## Next Step After Creating

Use `run_subagent` with the same filename to test:
```xml
<action>run_subagent</action>
<params>{"filename": "subagent.py", "query": "your test query"}</params>
```

## Core Component: call_llm

The LLM is the core reasoning engine of your subagent. Use it like this:

```python
from llm import call_llm

response = call_llm(
    system="You are a research assistant.",
    messages=[{"role": "user", "content": "Your prompt here"}],
    max_tokens=2000  # optional, default 8000
)
# Returns: str - The LLM's text response, or "Error:..." on failure
```

**IMPORTANT**: `call_llm` requires TWO positional arguments:
- `system`: A string for the system prompt
- `messages`: A list of message dicts with "role" and "content" keys

## Tool Skills in Subagent

**IMPORTANT**: Before using any tool skill, use `get_skill_description` to read its documentation and understand:
- How to import the tool
- What parameters it accepts
- What format it returns

Each tool skill has a SKILL.md file that shows exactly how to use it. For example:

```python
# For search tool - check its SKILL.md first with get_skill_description
from tools import local_search
results = local_search(query="your query", topk=10)

# For open_page tool - check its SKILL.md first with get_skill_description
from tools import open_page
content = open_page(docid="doc_id_from_search")
```

**Workflow**:
1. Meta agent calls `get_skill_description("local_search")` to read the tool's documentation
2. Meta agent creates subagent code that imports and uses the tool directly as shown in the documentation
3. Subagent runs and uses the tool by importing it directly

## Important Notes

- **CRITICAL: Data Passing Between Subagents**: When a subagent needs to collect large amounts of data, or when data needs to be passed between multiple subagents, use JSON files as the intermediate format. The producing subagent writes data to a `.json` file in the workspace, and the consuming subagent reads from that JSON file. This avoids passing large data through return values and keeps subagent interfaces clean.
- **CRITICAL: Output Files Must Use Relative Paths**: When the subagent needs to **write/save** output files (markdown, images, JSON, etc.), ALWAYS use **relative paths** (e.g., `output.md`, `result.png`, `data.json`). The subagent runs in a workspace directory, so relative paths will automatically save files to the correct location. Do NOT use absolute paths like `/Users/.../file.md` for output.
  - Examples of output files: `plt.savefig('chart.png')`, `pathlib.Path('result.md').write_text(...)`, `json.dump(data, open('data.json', 'w'))`
- **Reading Input Files**: For **reading** files (e.g., reading source code to analyze, reading data files), you CAN use absolute paths if needed — this is allowed since you're referencing existing files, not creating new output.
- **CRITICAL: Matplotlib Text Encoding**: When the subagent uses `matplotlib` for plotting, Do NOT use emoji characters (e.g., `\U0001f4c8`, `📈`, `😊`) in any labels, titles, legends, annotations, or text elements. Emojis are NOT supported by matplotlib's default fonts and will cause rendering failures or missing characters. Use plain text only.
- **CRITICAL**: Before creating a subagent, you MUST call `get_skill_description` for EVERY tool skill the subagent will use. This is mandatory to understand:
  - How to import the tool correctly
  - What parameters it accepts and their types
  - What format the tool returns (essential for writing correct code!)
- The subagent should be a **complete problem-solving pipeline**, not a simple function
- It should have its own reasoning loop: think → act → observe → iterate
- The subagent's `main(query)` function receives the question as a parameter and returns a dict
- Use `def main(query):` as the entry point
- The subagent should work for ANY question of the same category

## ⚠️ CRITICAL: Subagent MUST Be Generic ⚠️

A subagent will be saved and reused for MANY different questions of the same category. **NEVER hardcode the current question's specific details** (entity names, dates, search terms, expected answers, if-checks for specific strings) into the code.

From zzatpku/AgentFactory, MIT.

More from zzatpku/AgentFactory