GenZAgents for AutoGen — verified receipts for conversational agent groups

Conversational agent groups need turn-level capture

AutoGen is unique among the agent frameworks: agents talk to each other. The audit unit is the turn, not the tool call. Our wrapper captures every turn — the speaker, the message digest, the tool calls invoked in that turn — and rolls them up into one parent receipt covering the whole conversation. The dashboard renders this as a chat-style timeline so you can replay the agent dialogue.

Termination tracking

AutoGen conversations can terminate via several signals: max turns hit, termination message detected, exception. The wrapper captures the termination reason as a structured field. Knowing that conversation X terminated via "max_turns_exceeded" vs "termination_msg" is the difference between debugging a runaway loop and confirming successful completion.

Per-agent cost rollup

A typical AutoGen GroupChat has 3-4 agents on different models. The wrapper logs cost per agent per turn: at the end you see "UserProxy $0.00, Coder used GPT-5 $0.31, Reviewer used Claude $0.14". This makes the choice "which agent should we move to a cheaper model" data-driven.

AutoGen 0.4 actor-model support

AutoGen 0.4 switched to an actor model with explicit message passing. The wrapper supports both the v0.2 GroupChat API and the v0.4 RoutedAgent API — same receipt shape, same dashboard rendering. Use whichever AutoGen version your team is on; the integration covers both.

Streaming + interruption capture

AutoGen conversations can be interrupted by humans (UserProxyAgent prompts) or stop conditions. Each interruption is captured as a sub-event on the receipt — the audit log shows "user paused at turn 5, resumed at turn 6 with a corrected directive". Useful for human-in-the-loop workflows where the interruption pattern is itself the value.

Microsoft enterprise alignment

Microsoft customers running AutoGen typically also use Microsoft Entra (SSO), Azure OpenAI (LLM), and Azure Container Apps (hosting). GenZAgents is API-compatible with all three: Entra SSO for sign-in, Azure OpenAI for the embedding service, deployed on Azure Container Apps. The integration story for a Microsoft-aligned org is "everything stays in Azure, GenZAgents drops in alongside".

Install

$ npm install @genzagentsio/autogen

import { wrapGroupChat } from "@genzagentsio/autogen"
const groupChat = wrapGroupChat(rawGroupChat, {
  agentDid: "did:genz:...",
  apiKey: process.env.GENZAGENTS_API_KEY,
})
await groupChat.run(initialMessage)

What we capture

Every GroupChat.run: per-message agent attribution, tool calls, termination reason.

Verify it works

Run wrapped GroupChat → confirm one parent receipt with per-turn sub-events on /dashboard.

Common questions

Does the wrapper work with AutoGen Studio (the no-code UI)?▾

Indirectly. AutoGen Studio runs AutoGen behind the scenes; you need to modify the Studio backend to use our wrapper. We're working with the AutoGen team on a first-party integration; in the meantime, the manual wrap path is the supported route.

What about AutoGen's code-execution agents?▾

Captured. Code-execution turns include the executed code + the result (digest only by default), so you can audit what was run without storing the full output by default.

Can I capture only specific agents in a GroupChat?▾

Yes — pass `captureAgents: ["Coder", "Reviewer"]` to scope which agents' turns are captured. Other agents' turns are still included as parent metadata but no per-turn receipt event.

Does it work with AutoGen + Azure OpenAI?▾

Yes. The wrapper uses AutoGen's model_client interface; it doesn't care which provider sits behind it. Azure OpenAI works the same as direct OpenAI from the wrapper's perspective.