🧰 Skills

Skills are one of the two sources Agent Codemode uses to generate programmatic tools (the other being MCP Servers).

Skills are reusable code patterns that compose MCP tools to accomplish specific tasks. They allow agents to build up a toolbox of higher-level operations. Agent Codemode integrates with the agent-skills package for comprehensive skill management.

Skills APIs live in agent-skills. Import skill managers and helpers from agent_skills.

How It Works

Skills can be used in two complementary ways:

Pattern A: Direct Skill Toolset

The AgentSkillsToolset from agent-skills provides skill tools (list_skills, load_skill, run_skill_script, etc.) as a separate Pydantic AI toolset. This is the standalone pattern.

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────────────────┐
│  Skill Files    │ ──▶ │ Skills Manager  │ ──▶ │  Callable Python Functions  │
│  (skills/*.py)  │     │   (discovers    │     │   from skills.batch_process │
│   or saved via  │     │   and loads)    │     │   import batch_process      │
│     manager     │     │                 │     │                             │
└─────────────────┘     └─────────────────┘     └─────────────────────────────┘

Pattern B: Skills as Generated Bindings (Codemode Integration)

When skills are wired into Codemode, Agent Codemode generates typed Python bindings for skill operations — the same way it generates bindings for MCP server tools. This allows agents to call skill tools from within execute_code, using the same import pattern:

┌─────────────────┐     ┌─────────────────────┐     ┌─────────────────────────────────────┐
│  AgentSkills    │ ──▶ │ generate_skill_     │ ──▶ │  Generated Python Bindings          │
│  Toolset        │     │ bindings()          │     │   from generated.skills            │
│  (list, load,   │     │ (codegen.py)        │     │   import list_skills, run_skill      │
│   run, read)    │     │                     │     │                                     │
└─────────────────┘     └─────────────────────┘     └─────────────────────────────────────┘

This pattern is used by agent-runtimes when both Codemode and Skills are enabled. The skill tools are automatically wired into Codemode via wire_skills_into_codemode(), so agents can discover and call skill operations inside Python code alongside MCP tools.

Generated structure:

generated/
├── mcp/
│   ├── filesystem.py       # MCP server bindings
│   ├── web.py              # MCP server bindings
│   └── skills/             # Skill bindings (auto-generated)
│       ├── __init__.py
│       ├── list_skills.py
│       ├── load_skill.py
│       ├── read_skill_resource.py
│       └── run_skill.py

Usage in execute_code:

from generated.skills import list_skills, load_skill, run_skill

# Discover available skills
skills = await list_skills({})

# Load a skill's instructions
instructions = await load_skill({"skill_name": "pdf-extractor"})

# Run a skill script
result = await run_skill({
    "skill_name": "pdf-extractor",
    "script_name": "extract.py",
    "args": {"path": "/data/report.pdf"}
})

Lifecycle

1. Create - Write skill files in skills/ directory or save via agent_skills.SimpleSkillsManager
2. Discover - Agent Codemode discovers all available skills
3. Load - Skills are loaded as callable Python functions
4. Use - Import and call skills like any Python module (or via generated bindings in Codemode)

Overview

Skills enable AI agents to:

• Save useful patterns: Store code that works well for reuse
• Build a toolbox: Accumulate capabilities over time
• Compose operations: Chain multiple tools into higher-level functions
• Share knowledge: Export skills for use by other agents

Skills as Code Files

The primary pattern is organizing skills as Python files in a skills/ directory:

# skills/batch_process.py
"""Process all files in a directory."""

async def batch_process(input_dir: str, output_dir: str) -> dict:
    """Process all files in a directory.
    
    Args:
        input_dir: Input directory path.
        output_dir: Output directory path.
    
    Returns:
        Processing statistics.
    """
    from generated.mcp.filesystem import list_directory, read_file, write_file
    
    entries = await list_directory({"path": input_dir})
    processed = 0
    
    for entry in entries.get("entries", []):
        content = await read_file({"path": f"{input_dir}/{entry}"})
        # Process content...
        await write_file({"path": f"{output_dir}/{entry}", "content": content.upper()})
        processed += 1
    
    return {"processed": processed}

Skills as Standalone CLI Tools

Skills can also be designed as standalone command-line tools that can be run independently:

#!/usr/bin/env python
# skills/analyze_file.py
"""
Standalone CLI tool that can be run with: python skills/analyze_file.py <file-path>
This demonstrates a skill that can be invoked directly from the command line.
"""
import asyncio
import sys
import json

async def analyze_file(file_path: str) -> dict:
    """Analyze a file and return statistics."""
    from generated.mcp.filesystem import read_file
    
    content = await read_file({"path": file_path})
    lines = content.split('\n')
    word_count = len(content.split())
    char_count = len(content)
    
    return {
        "lines": len(lines),
        "words": word_count,
        "chars": char_count,
        "blank_lines": sum(1 for line in lines if line.strip() == '')
    }

# Run if called directly
if __name__ == "__main__":
    if len(sys.argv) < 2:
        print("Usage: python analyze_file.py <file-path>")
        sys.exit(1)
    
    file_path = sys.argv[1]
    result = asyncio.run(analyze_file(file_path))
    print(json.dumps(result, indent=2))

Run it directly:

python skills/analyze_file.py /data/myfile.txt
# Output: {"lines": 42, "words": 256, "chars": 1024, "blank_lines": 5}

Or import and use it in executed code:

from skills.analyze_file import analyze_file
stats = await analyze_file("/data/myfile.txt")
print(f"File has {stats['lines']} lines")

Helper Utilities as Skills

Skills can also provide reusable helper utilities for other skills:

# skills/helpers.py
"""Helper utilities for skill composition."""

async def wait_for(
    condition,
    interval_seconds: float = 1.0,
    timeout_seconds: float = float('inf')
) -> None:
    """Wait for an async condition to become true."""
    import time
    start_time = time.time()
    
    while True:
        result = await condition()
        if result:
            return
        
        if time.time() - start_time > timeout_seconds:
            raise TimeoutError(f"Timeout waiting for condition after {timeout_seconds}s")
        
        await asyncio.sleep(interval_seconds)


async def retry(fn, max_attempts: int = 3, delay_seconds: float = 1.0):
    """Retry a function until it succeeds or max attempts reached."""
    last_error = None
    
    for attempt in range(1, max_attempts + 1):
        try:
            return await fn()
        except Exception as e:
            last_error = e
            if attempt < max_attempts:
                await asyncio.sleep(delay_seconds)
    
    raise RuntimeError(f"Failed after {max_attempts} attempts: {last_error}")

Use these helpers in other skills:

from skills.helpers import retry, wait_for
from generated.mcp.web import fetch_url

# Retry a flaky API call
data = await retry(lambda: fetch_url({"url": "https://api.example.com/data"}))

# Wait for a condition
await wait_for(lambda: check_file_exists("/output/result.json"))

Using Skills in Executed Code

Skills are imported and called like any Python module within execute_code:

# In executed code
from skills.batch_process import batch_process

result = await batch_process("/data/input", "/data/output")
print(f"Processed {result['processed']} files")

SimpleSkillsManager

For programmatic skill management, use the SimpleSkillsManager from agent-skills:

from agent_skills import SimpleSkillsManager, SimpleSkill

# Create a skills manager
manager = SimpleSkillsManager("./skills")

# Save a skill
skill = SimpleSkill(
    name="backup_to_cloud",
    description="Backup files to cloud storage",
    code='''
async def backup_to_cloud(source: str, bucket: str) -> dict:
    files = await bash__ls({"path": source})
    uploaded = 0
    for f in files:
        await cloud__upload({"file": f, "bucket": bucket})
        uploaded += 1
    return {"uploaded": uploaded}
''',
    tools_used=["bash__ls", "cloud__upload"],
    tags=["backup", "cloud"],
)
manager.save_skill(skill)

# Load a skill later
loaded = manager.load_skill("backup_to_cloud")
print(loaded.description)  # "Backup files to cloud storage"
print(loaded.code)

SimpleSkill Attributes

Attribute	Type	Description
name	str	Unique skill identifier
description	str	Human-readable description
code	str	Python code implementing the skill
tools_used	list[str]	List of tool names used
tags	list[str]	Optional categorization tags
parameters	dict	Optional JSON schema for parameters
created_at	float	Unix timestamp when created
updated_at	float	Unix timestamp when last updated

SkillDirectory Pattern

For file-based skill organization with automatic discovery:

from agent_skills import SkillDirectory, setup_skills_directory

# Initialize skills directory
skills = setup_skills_directory("./workspace/skills")

# List available skills
for skill in skills.list():
    print(f"{skill.name}: {skill.description}")

# Get a specific skill
skill = skills.get("batch_process")

# Search for relevant skills
matches = skills.search("data processing")

# Create a new skill programmatically
skills.create(
    name="my_skill",
    code='async def my_skill(x: str) -> str: return x.upper()',
    description="Transform text to uppercase",
)

Composing Skills

Skills can import and use other skills to build higher-level operations:

# skills/analyze_and_report.py
"""Analyze data and generate a report."""

async def analyze_and_report(data_dir: str) -> dict:
    from skills.batch_process import batch_process
    from skills.generate_report import generate_report
    
    # First process the files
    process_result = await batch_process(data_dir, f"{data_dir}/processed")
    
    # Then generate a report
    report = await generate_report(f"{data_dir}/processed")
    
    return {"processed": process_result["processed"], "report": report}

MCP Server Integration

When running Agent Codemode as an MCP server, skills are exposed through dedicated tools:

• save_skill: Save a new skill or update an existing one
• run_skill: Execute a saved skill by name

from agent_codemode import codemode_server, configure_server
from agent_codemode import CodeModeConfig

config = CodeModeConfig(
    skills_path="./skills",
    # ... other config
)

configure_server(config=config)
codemode_server.run()

Pydantic AI Integration

For Pydantic AI agents, use the AgentSkillsToolset:

from pydantic_ai import Agent
from agent_skills import AgentSkillsToolset, SandboxExecutor
from code_sandboxes import LocalEvalSandbox

# Create toolset with sandbox execution
sandbox = LocalEvalSandbox()
toolset = AgentSkillsToolset(
    directories=["./skills"],
    executor=SandboxExecutor(sandbox),
)

# Use with pydantic-ai agent
agent = Agent(
    model='openai:gpt-4o',
    toolsets=[toolset],
)

# Agent gets: list_skills, load_skill, read_skill_resource, run_skill_script

Helper Utilities

The agent-skills package provides helper utilities for skill composition:

from agent_skills import wait_for, retry, run_with_timeout, parallel, RateLimiter

# Retry an operation on failure
result = await retry(my_async_function, max_attempts=3, delay=1.0)

# Run with timeout
result = await run_with_timeout(slow_operation(), timeout=30.0)

# Run multiple operations in parallel
results = await parallel([task1(), task2(), task3()])

# Rate limiting
limiter = RateLimiter(max_calls=10, period=60)  # 10 calls per minute
async with limiter:
    await api_call()

Best Practices

1. Clear naming: Use descriptive names that indicate what the skill does
2. Documentation: Include docstrings with Args, Returns, and Examples
3. Single responsibility: Each skill should do one thing well
4. Error handling: Include try/except for robust execution
5. Composability: Design skills to be easily combined with others
6. Testing: Test skills independently before composing them
7. Codemode integration: When both are enabled, prefer using skill bindings via generated.skills for consistency with MCP tool patterns

Generated Skill Bindings (Approach D)

When Codemode and Skills are both enabled in agent-runtimes, skills are automatically wired into Codemode through a process called Approach D: Skills as Generated Bindings.

How It Works

1. Binding generation: generate_skill_bindings() in codegen.py produces async Python wrapper functions for each skill tool (list_skills, load_skill, read_skill_resource, run_skill).

2. Executor routing: When code in the sandbox calls await list_skills({}), the call is routed through the executor's call_tool("skills__list_skills", args), which delegates to a skill tool caller callback connected to the real AgentSkillsToolset.

3. Post-init callback: Since Codemode lazily initializes its executor, skill wiring uses a post-init callback (add_post_init_callback) to defer binding generation and caller registration until the executor is ready.

Call Flow

Sandbox Code                    Executor                       AgentSkillsToolset
     │                              │                                  │
     │  from generated.mcp.     │                                  │
     │  skills import run_skill     │                                  │
     │                              │                                  │
     │  await run_skill({...})      │                                  │
     │ ─────────────────────────▶   │                                  │
     │                              │  call_tool("skills__run_skill")  │
     │                              │ ─────────────────────────────▶   │
     │                              │                                  │  execute skill
     │                              │     ◀────────────────────────    │
     │     ◀────────────────────    │                                  │
     │         result               │                                  │

Remote Sandbox Support

When using a remote sandbox (e.g., Datalayer Runtime or Jupyter kernel), skill bindings are generated inline inside _generate_tools_in_sandbox() alongside MCP tool bindings. The sandbox calls back to agent-runtimes via HTTP, where a skills proxy routes the request to the real AgentSkillsToolset:

Remote Sandbox                   agent-runtimes                    AgentSkillsToolset
     │                              │                                  │
     │  await run_skill({...})      │                                  │
     │ ──── HTTP POST ────────▶     │                                  │
     │    /mcp-proxy/skills/        │                                  │
     │    run_skill                  │  skills_proxy_caller(...)        │
     │                              │ ─────────────────────────────▶   │
     │                              │                                  │  execute skill
     │                              │     ◀────────────────────────    │
     │     ◀────────────────────    │                                  │
     │         result               │                                  │

Post-Init Callbacks

The CodemodeToolset supports deferred initialization callbacks via add_post_init_callback(). This is used by agent-runtimes to wire skills into Codemode at the right time:

from agent_codemode import CodemodeToolset

toolset = CodemodeToolset(registry)

# Register a callback that fires once after lazy initialization
toolset.add_post_init_callback(lambda ts: wire_skills(ts))

# Callbacks fire automatically when the executor is first initialized
# (e.g., on the first call to execute_code)

See Also

• Agent Skills Documentation
• Programmatic Tools
• Tools Discovery