AI Engineer & Researcher. I am currently an AI Researcher at Vatic Labs.
My work spans hierarchical multi-agent orchestration, cross-environment agent learning, scalable environment generation, agent workflow optimization, and protocols for agentic societies. I co-founded OpenManus, an open-source agent infrastructure with 56k+ GitHub stars, and work with the Foundation Agents community. I received my B.E. from Renmin University of China.
Focus
- Agent learning. Learning is key to cross-environment capabilities. Learning environment dynamics requires complex optimization approaches, signals, and targets beyond model parameters, like AFlow optimizing decision workflows and SPO exploring new reward signals for prompt optimization.
- Decision-making. Human decision-making naturally enables cross-environment learning and generalization. AOrchestra automates sub-agent creation, AoT atomizes reasoning to address context limitations, and ReCode unifies planning and action for more natural decision-making.
- Environment scaling. Agent environments are simplified versions of human environments, lacking complexity, dynamics, and rich reward signals. We explore scalable environments through AutoWebWorld, Scalable Environments, and AutoEnv.
- Agent society. As agent systems become interoperable infrastructure, Foundation Protocol studies coordination layers for multi-agent organization, value exchange, provenance, and governance.
Selected Publications
-
ReCode: Unify Plan and Action for Universal Granularity Control.
-
AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration.
-
AutoWebWorld: Synthesizing Infinite Verifiable Web Environments via Finite State Machines.
[paper]
-
InteractComp: Evaluating Search Agents With Ambiguous Queries.
-
VisJudge-Bench: Aesthetics and Quality Assessment of Visualizations.
-
Foundation Protocol: A Coordination Layer for Agentic Society.
-
Scalable Environments Drive Generalizable Agents.
[paper]
-
AutoEnv: Automated Environments for Measuring Cross-Environment Agent Learning.
-
AFlow: Automating Agentic Workflow Generation.
-
Self-Supervised Prompt Optimization.
-
Atom of Thoughts for Markov LLM Test-Time Scaling.
-
Reasoning via Video: The First Evaluation of Video Models' Reasoning Abilities through Maze-Solving Tasks.
-
Advances and Challenges in Foundation Agents: From Brain-Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems.
-
MobileExperts: A Dynamic Tool-Enabled Agent Team in Mobile Devices.
[paper]
Selected Projects
-
No fortress, purely open ground. OpenManus is Coming.
-
Next paradigm for LLM Agent. Unify plan and action through recursive code generation for adaptive, human-like decision-making.
-
About Awesome things towards foundation agents. Papers / Repos / Blogs / ...
-
🔥🔥🔥 ICLR 2025 Oral. Automating Agentic Workflow Generation.
Experience
-
AI Researcher, Vatic Labs
Mar 2026 – Present -
Researcher, DeepWisdom
Feb 2025 – Mar 2026 · Shenzhen
Co-founded OpenManus and advanced Foundation Agents research with Chenglin Wu, Jiayi Zhang, and Yifan Wu. -
Algorithm Engineer Intern, Xiaomi
Jan 2024 – Apr 2024 · Beijing
with MiRoboticsLab -
Research Assistant, GeWu Lab (Renmin Univ)
Jun 2022 – Aug 2023 · Beijing
with Prof. Di Hu
Service
- Reviewer: ICML 2026; ICLR 2026; CVPR 2026; ECCV 2026; ICML 2025 MAS Workshop
Talks
-
“Advances and Challenges in Foundation Agents”
Invited talk at 2025 X-AGI & The 18th China-R Conference, Beijing (Oct 2025)
Blogs
-
规划与行动统一:ReCode 对 Agent 决策的重新思考 2025.11.03 · Chinese
用代码统一表示 plan 和 action,结合动态展开机制,可能为 foundation agent 的 learning 提供一个更好的 decision-making 基础,但这只是一个初步探索。
-
【小米CyberDog二次开发】让你的机器狗有自己的小情绪! 2022.09.14 · Chinese
基于小米CyberDog和文澜预训练模型,我们实现了铁蛋根据不同场景而做出不同的富有情绪动作的能力。下面是保姆级教程(基于Python实现),无需ROS相关知识,让你的铁蛋也拥有自己的情绪!