AutoGen Integration

Strathon enforces policies on AutoGen tool calls before they execute, across multi-agent conversations, with the full action set including interactive approval. Integration wraps BaseChatAgent.on_messages and captures conversations, tool calls, and LLM interactions.

Enforcement scope: full. The pre-execution hook is async, so all seven actions enforce: block and throttle stop the call, steer substitutes the tool result directly, and require_approval pauses until an operator decides (and fails closed on expiry).

Installation

bash
pip install "strathon[autogen]"

Install this in your agent's environment, not alongside the receiver. autogen-core requires protobuf~=5.29.3 and the receiver requires protobuf>=7.35.1. Those ranges do not overlap, so pip resolves one and reports the other as unsatisfied. The two are designed to run apart — the receiver in its own container or virtualenv, the SDK inside the agent application — so this only appears if both land in one environment. If you are testing the AutoGen adapter, give it a clean virtualenv; an adapter exercised against an unsupported protobuf is not a trustworthy result.

Setup

python
from strathon import Client, instrument

client = Client(
    api_key="stra_...",
    endpoint="http://localhost:4318",
)
instrument(client, frameworks=["autogen"])

What Gets Captured

  • Agent messages: sender, receiver, content
  • Tool calls: function name, arguments, return value
  • Multi-agent conversations: full message thread
  • LLM calls: model, tokens, latency

Example Policy

Block code execution tools outright:

cel
attrs["gen_ai.tool.name"] == "execute_code"

Block code execution in production (AutoGen group chats re-send the full history every turn, so a runaway loop is expensive: gate the dangerous tool):

cel
attrs["gen_ai.tool.name"] == "execute_code"
  && attrs["strathon.project.environment"] == "production"

Notes

  • Wraps BaseChatAgent.on_messages and BaseGroupChat.run for conversation and team tracing, and BaseTool.run_json to enforce policies on each tool call (the enforcement boundary). All installed at instrument time.
  • Requires autogen-agentchat>=0.7.5 (installed by the autogen extra).
  • Multi-agent conversations create a single trace with per-agent spans.

Learn More