Paper Review | Autogen

Enabling Next-Gen LLM Applications via Multi-Agent Conversation Framework

One of my problems when building LLM-based solutions is that Langchain is really weird to work with since it’s changing quietly on so many corners. However, after meeting Yann LeCun (see my recent Note) the other day, I am more than ever convinced that the direction I am going is the right one. Clearly, frameworks for developing LLMs are needed. And, AutoGen IS a framework for developing LLM applications using multiple agents that can converse with each other to solve tasks. So, given the parsing problems I had with Langchain, let’s explore what Autogen can do for me by starting with the paper.

Project Goal: Establish a framework for the robust development of LLMs

Problem: While LLMs like my ReAct Agent with Memory are powerful, most standard LLMs can’t reason well, have no memory, and don’t have a world model past their training data.

Proposed Solution: Developed and open-sourced by Microsoft Research, Autogen is a multi-agent system where LLMs converse with each other to solve a task. So key features include customizable agents that are easy to integrate with humans and tools, a standardized conversation interface, and flexible conversation patterns in complex workflows.

Opinion: Microsoft has introduced the concept of workflows as a sequence of conversations. I think that makes a lot of sense. In addition, Microsoft has the engineering track record to establish a developer-friendly solution where especially enterprise architects can trust upon when designing their solutions. Since Microsoft has open-sourced their framework we can all get involved. The agents appear to be quite powerful, I will get back to this in my code clinic.

Links: Paper, Github, Website, Discord (7/10)

Let’s dive in.

As someone who has led large distributed teams, getting everyone aligned into an efficient workflow is really hard. So how can we achieve this for a workflow of LLMs, Humans, and Tools? Microsoft research proposes a framework that brings all of this into a sequence of conversations and calls it Autogen. Autogen has three main features (1) Agents, (2) Messaging Standards, (3) Sequence Patterns.