agent-framework-ep

Microsoft Agent Framework extensions for mainstream open-source LLMs, including structured output support for GLM, Kimi, Qwen, and DeepSeek, along with reasoning_content support. Plus a local containerized code interpreter environment. 'ep' stands for enterprise-level applications.

Features

• OpenAI-like Client Extensions (openai_like)

  • Structured output parsing with JSON fallback (dirtyjson, json-repair)
  • Reasoning content support for DeepSeek-R1 style models
  • Uses Chat Completions API, compatible with domestic LLMs (DeepSeek, Kimi, Qwen)
  • Requires: agent-framework>=1.0.0rc6 and agent-framework-openai>=1.0.0rc6

• Code Executor (code_executor)

  • Docker-based code execution environment
  • Supports Python, bash, and shell scripts
  • Timeout and cancellation support
  • Isolated execution for security

• Dynamic Skills Provider (skills_provider)

  • Async skill updates before each agent run
  • Extendable skills from external sources

Installation

pip install agent-framework-ep

Or with uv:

uv add agent-framework-ep

Prerequisites

• Python 3.12+
• Docker (for code execution features)

Quick Start

OpenAI-like Client with Structured Output

from pydantic import BaseModel
from agent_framework import Agent
from agent_framework_ep import OpenAILikeChatClient

class Response(BaseModel):
    answer: str
    confidence: float

# Create client with structured output support
client = OpenAILikeChatClient(
    model="deepseek-chat",
    api_key="your-api-key"
)

# Use with Agent framework
agent = Agent(client=client)
response = await agent.run(
    "What is the capital of France?",
    response_format=Response
)
print(response)  # Parsed Response object

Code Execution

import asyncio
import os
from agent_framework_ep import DockerCommandLineCodeExecutor
from agent_framework_ep import OpenAILikeChatClient, CodeExecutionTool

chat_client = OpenAILikeChatClient(
    model_id="kimi/kimi-k2.5",
)

code_executor = DockerCommandLineCodeExecutor(
    image="python-code-sandbox",
    work_dir="some_of_your_working_directory",
    delete_tmp_files=True,
    environment={
        "TAVILY_API_KEY": os.environ.get("TAVILY_API_KEY"),
    }
)

code_tool = CodeExecutionTool(code_executor).execute_code

agent = chat_client.as_agent(
    name="SkillsAssistant",
    instructions="You're a helpful assistant.",
    tools=[code_tool],
)

Dynamic Skills Provider

from agent_framework_ep import UpdatableSkillsProvider
from agent_framework import Skill

async def fetch_dynamic_skills():
    # Fetch skills from external source
    return [
        Skill(name="web-search", description="Search the web", content="..."),
    ]

provider = UpdatableSkillsProvider(
    skill_paths="./skills",
    skills_updater=fetch_dynamic_skills
)

Development

# Clone the repository
git clone https://github.com/qianpeng/agent-framework-ep.git
cd agent-framework-ep

# Install dependencies
uv sync

# Run tests
uv run pytest

# Run tests with Docker (requires Docker)
uv run pytest -m docker

# Lint and format
uv run ruff check --fix .
uv run ruff format .

# Type check
uv run mypy src/agent_framework_ep

License

MIT License - see LICENSE for details.