OpenBMB Meetup recap: the team behind a 28K-star project on how multi-agent systems are reshaping human-AI collaboration
On the evening of January 17, the OpenBMB open-source community held a meetup themed “AI Multi-Agent Creative Workflow Architecture” at AGI Bar · Knowledge Distillation. The event invited Associate Professor Qian Chen from Shanghai Jiao Tong University and the core team of the ChatDev 2.0 project he leads to share and discuss the latest research on multi-agent systems, real-world applications, and future outlook.
The session drew more than 60 registered participants. The project team systematically traced the evolution of agent technology from theoretical exploration to practical application, sketching a clear development path and implementation blueprint. More than a solid technical exchange, the event was a collision of ideas and a spark of inspiration about the future of human-AI collaboration 💥
Opening: the OpenBMB community and ChatDev 2.0
To kick off the event, Jing Chenzhe, head of the OpenBMB open-source community, walked developers through the community’s origins and development. OpenBMB was jointly initiated by THUNLP and ModelBest, and has always held to the core principle of “technology for all, open to all.” It works to build a complete open-source ecosystem — from the lightweight, efficient foundation model MiniCPM and the advanced multimodal model MiniCPM-V to the speech generation project VoxCPM and development frameworks such as BMTrain and the retrieval system UltraRAG — to make the vision of “bringing large models to every home” a reality. Jing then spotlighted the community’s flagship project, ChatDev 2.0, whose frontier exploration of multi-agent collaboration turned the audience’s attention to the future. A deep discussion of agent evolution and ecosystem integration was underway 🔛
Talk: Qian Chen on multi-agent collaboration and evolution in large models 🧬: Qian Chen, Associate Professor at Shanghai Jiao Tong University, postdoc in the Department of Computer Science and Technology at Tsinghua University, and head of ChatDev 2.0 R&D, together with Dang Yufan from Tsinghua University and Yao Shu from Shanghai Jiao Tong University, shared their work on multi-agent collaboration and evolution in large models.
As data growth levels off, the industry is beginning to realize that a Scaling Law based purely on growing model parameters may hit a bottleneck between 2026 and 2028. “When a single model’s capacity to store intelligence has a theoretical ceiling, the path to artificial general intelligence may require a change of approach,” Professor Qian said. The future, he suggested, may lie in “spreading intelligence across multiple models” and combining them cleverly into a more powerful intelligence system. Behind this idea are the two core goals of multi-agent systems: decomposing and coordinating capabilities, and enabling efficient human-AI collaboration. Multi-agent systems can break complex capabilities into parts and achieve scalability through coordination at a higher layer, overcoming the limits of a single model that “does everything.” That means humans only need to give one high-level instruction, and the agents — like a well-trained team — divide the work and collaborate on their own to complete tasks ranging from game development to web design.
This multi-agent philosophy comes to life in the ChatDev project. ChatDev 1.0 took an imaginative technical approach: simulating the workflow of a software company. In this system, the human user plays the “boss,” while agents take on roles such as CEO, CTO, programmer, and tester. ChatDev 2.0 represents a qualitative leap, with a core philosophy built around code. Version 2.0 can handle downstream tasks mediated by code, such as data analysis, text reports, and even 3D model and video generation. It also dramatically lowers the barrier to building agents: users can assemble their own agent systems through a visual interface or simple drag-and-drop.
Efficient multi-agent systems depend on continued breakthroughs at the infrastructure layer. The team explored agent learning in depth, letting agents accumulate experience from past tasks. “Around 2024, we began exploring how agents could learn from execution traces and reuse that experience in future reasoning, significantly improving task completion efficiency,” a team member said. For multi-agent coordination, the team adopted a distinctive “marionette” approach 🎭. “We train a centralized ‘super robot’ to coordinate with context awareness.” This central coordinator can call on different agents based on task progress and real-time state. The related results have been published at top conferences including ACL and NeurIPS.
Live Q&A: all about ChatDev 2.0 🎙️
During the Q&A, attendees had a lively discussion about the practical challenges facing multi-agent systems.
Q: How do you balance controllability and creativity in agents?
A: Generative models are inherently creative, but that comes at the cost of some stability. We strengthen controllability at the knowledge and process levels through external tools (such as RAG retrieval) and workflow orchestration. Going forward we will keep exploring “process controllability” and “tool controllability,” but creativity and stability still require trade-offs depending on the scenario.
Q: In education, can AI be used for personalized student learning?
A: Technically yes, but we currently focus on the teacher side. Giving students direct control would require considering whether they can judge the correctness of AI-generated content. A “teacher review” step could be introduced, but we also need to watch the review burden on teachers. For general, well-structured courses, AI-generated content is already quite reliable.
Q: How is ChatDev different from other agent frameworks such as LangGraph?
A: ChatDev’s core positioning is a drag-and-drop, low-code multi-agent workflow system. The main difference from frameworks like LangGraph is that ChatDev aims to build workflows through intuitive drag-and-drop operations without writing large amounts of complex code. That makes it a low-cost tool for rapid trial and error and idea validation. We do not claim to be more stable than LangGraph or LangChain, but our implementation is simpler and easier to pick up. It is well suited to researchers prototyping multi-agent systems quickly.
Q: How do multi-agent systems control cost and improve efficiency?
A: We optimize agent collaboration paths to avoid redundant calls. A centralized orchestration system (such as the Puppeteer model) can dynamically select the most suitable agent for a subtask, improving results while significantly reducing token consumption. In the future we will further support “cost-aware” mechanisms that let users set call limits.
Q: As data becomes scarce, how can agents keep evolving?
A: Data growth may indeed slow, but we believe there is a theoretical paradox in using AI-generated data to retrain AI. A more viable path is multi-agent collaboration and social simulation, achieving capability emergence and task generalization with limited data.
Event spotlight: hands-on fun with ChatDev 2.0
The event was about more than talks — it was about seeing things work firsthand. Both the 🎁 raffle system and the 🖨️ AGI fortune-telling photo printing system were built on the ChatDev 2.0 agent platform ✅. The raffle system automatically generated the full raffle logic and interactive interface from the prize list and rules, and ran all the on-site interactions smoothly and reliably. From idea to running program was just one step — a real microcosm of multi-agent systems in production.
The AGI fortune-telling photo printing system closed the loop from generation to printing, a complete “AGI identity” system. Based on a photo taken on site, it generated the viewer’s virtual identity and exclusive avatar in a future AGI scenario in real time (for example, “Architect of Universal Balance”), then drove a photo printer to produce a physical print on the spot. More than an interaction, it was a personalized AI souvenir to take home.
🙌 The event fully reflected the OpenBMB community’s practical spirit — “from theory to real-world deployment, from solo development to swarm intelligence.” We believe that whether you are a researcher focused on the frontier, an engineer eager to build, or a student curious about the future, you found resonance, inspiration, and room to collaborate. As an open, open-source multi-agent development platform, ChatDev 2.0 always welcomes all developers, teams, and organizations interested in agent technology to join us in exploring more possibilities for human-AI collaboration and shaping the future ecosystem of agent applications 🚀
The OpenBMB community will keep bringing developers around the world more efficient, more intuitive, and more creative development experiences and opportunities to connect. See you next time ✨!
➤ Repository
🔗 https://github.com/OpenBMB/ChatDev
If our projects interest you ✨, scan the QR code to join the group, read the detailed developer docs, and talk directly with the core team!