AutoGen and AG2: Chi Wang and Qingyun Wu on Building Multi-Agent Systems

Matthew Berman interviews AutoGen creators Chi Wang, then a senior staff researcher at Google DeepMind, and Qingyun Wu, an assistant professor at Penn State. They trace the framework’s origins, discuss open source and practical agent design, and explain why they continued their work in AG2.

Wu says their collaboration began with FLAML, an open-source library for automated machine learning and hyperparameter tuning. She began working with Wang during a Microsoft Research internship in 2019. Their efficient tuning research became a library used by businesses, including banks working with tabular data.

Seeing the potential of generative models, they asked how to help people use them effectively. Wang recalls applying AutoML to GPT-3 inference parameters: choices could materially affect both performance and cost on mathematics, coding and text generation.

Wang says useful applications often involve more than a single model call: tools, multiple inference steps and human input form a large design space. They wanted a common framework that would let developers navigate that space and eventually optimize it, much as established machine-learning frameworks provide common interfaces.

Berman asks when they first recognized the potential of agents collaborating. Wu recalls choosing challenging math problems because they were important, relatively easy to evaluate and relevant to her interest in tutoring. Solving them often required writing, executing and debugging code, as well as a way to involve a human tutor.

That led to a user proxy agent alongside the assistant. Wu explains that, at the time, a user might copy code generated by ChatGPT, run it, and return the result; the proxy could perform that role and also accommodate human feedback. Watching the two agents iterate and debug autonomously was an early moment of excitement for her.

Wang connects his own insight to ChatGPT’s conversational interface, which let people give feedback and iterate. A lesson from a college seminar on dialogue as a way to create knowledge came back to him: capable language models might converse with one another, with humans participating when appropriate.

He saw different roles for reasoning, planning, coding, testing and review, enabling a larger workflow than one model call. For example, an agent simulating different users might help generate training data. He saw a broader opportunity to model complicated processes as conversations.

Their initial math case study paired a model-backed assistant with a proxy that could execute code or obtain manual feedback. Wang says early users then wanted more roles and communication patterns, prompting the move from two agents to larger multi-agent conversations.

Asked about the timeline, Wang says their thinking began in late 2022. Around the GPT-4 release in 2023, they began building around agents as the abstraction, iterating with users and writing a research paper that August. Wu adds that the first code was public inside FLAML from the start, although the wider audience initially did not notice it.

Wu says she was surprised by the popularity and motivated to improve the project. Wang had set a one-year target of 1,000 GitHub stars based on FLAML’s gradual growth, but says AutoGen exceeded his expectations within days. Berman notes that the project had over 35,000 stars at the time of the interview.

Berman asks how they came to open source. Wang says his earlier Microsoft Research work mostly produced papers and occasional conversations with product teams. FLAML began as research too; when people asked to use the code, the team initially had a prototype and offered a private license rather than an open-source release.

When another reader asked for open-source code, Wang decided they should release it despite uncertainty about adoption. Addressing user concerns improved usability and generated follow-up research. In his account, adoption grew from smaller companies to large organizations, including Microsoft products.

That experience changed their approach to AutoGen: code was open from the beginning to get feedback and iterate quickly. Wang believes this accelerated rather than impeded enterprise adoption and gave the team insight into many enterprise use cases.

Wu describes three stages in her relationship with open source. As a PhD student, sharing code was a default linked to reproducibility and making work accessible. With FLAML, seeing people apply her research at scale brought personal satisfaction; AutoGen led her to think more about the ecosystem’s value to contributors and users.

Wu argues that open code lets contributors build on one another’s work and gives enterprises transparency. A widely used ecosystem can also make onboarding easier because employees may already know the infrastructure. Berman values being able to inspect, modify and contribute improvements himself.

Turning to practical advice, Berman asks about model choice, agent count and prompts. Wu recommends starting by prototyping one working instance of a task. A small successful example is rewarding and lets a developer compare approaches before scaling the design.

Wang says the initial AutoGen guidance was to start simply: put the instructions in an assistant agent and give a user proxy appropriate capabilities, such as code execution when needed. Two agents may be enough for a simple task. As tasks become complex, a single model may forget instructions or perform some steps poorly.

At that point, Wang suggests splitting the work into smaller pieces and giving specialized agents narrower responsibilities. Developers must also decide how those agents communicate: with three or more participants, many sequences are possible, so the conversation topology becomes a design choice.

He describes sequential chats, nested chats and group chats as building blocks. They offer different tradeoffs between flexibility, control and isolation of an agent’s internal work. Different agents can also use different models, configurations and tools, allowing developers to balance quality, cost and latency.

Wang then describes two more recent AG2 features as of the interview: CaptainAgent and Swarm. CaptainAgent can analyze a supplied task, decompose it and assemble agent teams, drawing on existing agents and tools or generating new ones. The user can inspect what was created.

For difficult tasks or missing tools, Wang cautions that automatic creation may fail to produce everything needed. Developers can take useful parts as a starting point, then add missing tools or specialist roles. Swarm provides an interface for coordinating agents and specifying handoffs, shared context and structured output; he suggests combining the approaches.

Wu offers a general heuristic: look at how human organizations divide a problem. For instance, an HR department has different functional roles. She treats this as a starting analogy for organizing model-based agents, not a guarantee that a human organizational chart will produce the best system.

Berman asks when humans should be in the loop. Wang approaches the question in two stages: what human involvement remains desirable even with much stronger agents, and what additional work is necessary while the technology remains imperfect.

First, Wang says humans should supply the original intent and decide whether to accept suggestions. Since a brief request can be ambiguous, they may need to clarify that intent through several interactions. After an agent returns results, a person may need to verify them and provide feedback.

A fourth role is overriding an agent’s decisions or taking over when necessary. Wang says the degree of involvement in each role may change as capabilities improve. He also describes teaching agents missing knowledge and skills, and conversely learning from their work out of curiosity, for verification, or to develop one’s own abilities.

Wu suggests looking for gaps between models and humans when choosing where to involve people. Particularly challenging work may require human expertise and verification, while ethical and security considerations matter in sensitive applications. These are places she says human participation deserves serious consideration.

Berman asks why the founders forked the original AutoGen code into AG2. Wu says AutoGen began as a collaboration between Microsoft and outside contributors, and they initially agreed to publish it under Microsoft with an MIT license. For more than a year, contributors from multiple institutions worked together there.

As the project grew, Wu says organizational priorities and management made progress slower. She describes the desire to make decisions and move quickly with the community, while emphasizing that the constraints were a reality of a large company rather than simply a complaint about it.

Wu gives a second reason: a neutral home could make participation easier for contributors who might hesitate to work under another major company’s organization. At the time, she says, AG2 had contributors from Google, Meta, Microsoft, IBM, smaller companies and universities.

Asked about plans, Wu says development should remain community- and use-case-driven, with feedback from users guiding priorities. She cites recent work on CaptainAgent, Swarm, GraphRAG and structured output, and identifies better support for real-time use cases such as voice agents as a near-term goal in 2024.

Berman asks when agents might proactively work for people around the clock. Wu thinks the technology is close in some respects, but emphasizes interaction design and authorization. An agent might draft an important email, for example, while leaving the user to review it before sending.

Wang agrees that in specific domains there may be no fundamental technical barrier to long-running tasks. Another agent could even initiate requests. But this assumes sufficient computing resources and raises cost concerns; more importantly, quality is not guaranteed and an agent could eventually take an unwanted action.

Wang compares the difficulty of allocating responsibility to questions raised by autonomous driving. Beyond technical feasibility and regulation, he asks whether people actually want to delegate everything. His forecast is that in five to ten years agents may complete most tasks from a human idea with minimal supervision, but he stresses that this is an outlook, not an established capability.

For Wang, people should retain the choice to delegate, act themselves or take over. If capable agents become widespread, deciding what to do personally may itself become an important question. Berman closes by saying he would value agents that prepare work, such as an email draft, for his own review.