AutoGen

Verified 2026-08-19 — PASS, 5s. Session persistence: save_state() / load_state() blob.

AutoGen serialises the agent’s whole conversation state, so resume keeps structured turns rather than replayed text.

---

The runner

import sys, os, json, asyncio, pathlib
sys.path.insert(0, os.path.expanduser("~/agenttests"))
import azcfg, hetool

from autogen_agentchat.agents import AssistantAgent
from autogen_ext.models.openai import OpenAIChatCompletionClient

STATE_DIR = pathlib.Path(os.path.expanduser("~/agenttests/autogen/state"))
STATE_DIR.mkdir(parents=True, exist_ok=True)

def get_working_folder() -> str:
    """Return the agent's current working folder (absolute path)."""
    return hetool.get_working_folder()

def run_shell(command: str) -> str:
    """Run a shell command in the working folder and return its output."""
    return hetool.run_shell(command)

async def invoke(sid, message):
    c = azcfg.load()
    model_client = OpenAIChatCompletionClient(
        model=c["model"], base_url=c["base_url"], api_key=c["api_key"],
        model_info={"vision": False, "function_calling": True, "json_output": True,
                    "family": "unknown", "structured_output": True},
    )
    agent = AssistantAgent("assistant", model_client=model_client,
                           system_message="Be concise. Use tools for operational questions.",
                           tools=[get_working_folder, run_shell],
                           reflect_on_tool_use=True)

    f = STATE_DIR / f"{sid}.json"
    if f.exists():
        await agent.load_state(json.loads(f.read_text()))

    result = await agent.run(task=message)
    reply = result.messages[-1].content.strip()

    f.write_text(json.dumps(await agent.save_state()))
    await model_client.close()
    return reply

Async — the adapter drives the coroutine, so invoke may be async def.

---

Engine definition

Same shape as every other framework; only file and the runner path change.

{
  "type": "process",
  "file": "/home/you/autogen/.venv/bin/python3",
  "nativeSession": true,
  "argsNew":    ["/home/you/he_adapter.py", "/home/you/autogen_session.py", "",            "{message}"],
  "argsResume": ["/home/you/he_adapter.py", "/home/you/autogen_session.py", "{sessionId}", "{message}"],
  "replyField": "result",
  "sessionIdField": "session_id",
  "approve": false,
  "workdir": "/home/you/work",
  "proxy": { "shape": "openai", "baseUrlEnv": "OPENAI_BASE_URL", "keyEnv": "OPENAI_API_KEY",
             "modelEnv": "OPENAI_MODEL", "model": "<route>|<provider model>" }
}

---

Adoption

1. Keep your AssistantAgent and tools as they are.
2. Wrap agent.run(...) in async def invoke(sid, message).
3. Add load_state / save_state keyed on sid.
4. Build OpenAIChatCompletionClient from azcfg.load().
5. Seal the engine, point external.engine at it, restart.

---

Gotchas

model_info is required for any non-OpenAI model. The client cannot infer capabilities for an unknown model id and will refuse to start without it. function_calling: True is the field that matters if you bind tools.

reflect_on_tool_use=True or you get the raw tool output. Without it the agent’s last message is the tool result, not a composed answer — so result.messages[-1].content is JSON rather than prose.

Close the client. await model_client.close() before returning, or the process can hang with the connection pool open — which surfaces as a turn that never completes.

Plain callables are fine as tools. No decorator needed; AutoGen introspects the signature and docstring, so write a real docstring — it becomes the tool description the model sees.