Audit trail integration for AG2 agents

[jagmarques]

Running AG2 agents in production means you often need a record of what each agent decided and why. Regulators and enterprise compliance teams are increasingly asking for this, and stitching it together manually from logs is painful.

I built Asqav to solve this. It captures agent reasoning, tool calls, input/output at each step, and stores structured audit trails you can query or export. It works via a decorator pattern so it sits outside your agent logic.

For AG2, the natural integration point would be wrapping ConversableAgent.initiate_chat or hooking into the reply function chain. Something like:

import asqav
from autogen import ConversableAgent

audit = asqav.Audit(project="my-ag2-app")

@audit.trace
def run_chat():
    agent.initiate_chat(other_agent, message="...")

I can build a tighter AG2 integration if there is interest. Would an asqav-ag2 contrib package or a native hook point in AG2 make more sense? Happy to submit a PR either way.