ByteDance and Li Hang: Making Neural Networks Work Instead of Just Chat

Dr. Li Hang, head of ByteDance's AI lab, has published a paper on a general-purpose framework for AI agents built on four modules: perception, planning, memory, and action. According to Jiqizhixin (机器之心), the goal is to replace today's experimental agent systems such as AutoGPT and BabyAGI, which often "break on the second step and hallucinate in endless loops," with a reliable industrial-grade tool.

Who is Li Hang

In natural language processing, Li Hang's name carries, as the article puts it, "roughly as much weight as the entire server fleet of an average startup." His career ran through Microsoft Research Asia and Huawei before he took charge of ByteDance's AI lab. That is an applied track, not an academic one. In the author's view, this background suggests Li Hang cares less about the theoretical elegance of architectures than about their practical power — moving ideas out of PDFs and into real products.

Four architectural nodes

The framework targets a core weakness of today's language models: they can't plan ahead. A model "lives in the moment," generating the next token rather than building a strategy. The proposed architecture focuses on four critical nodes: perception, planning, memory, and action. The model first builds a hierarchical tree of goals, then uses external tools to reach them while continuously checking against long-term memory. The author sums up the effect this way: the approach turns AI "from a philology student into a seasoned project manager."

Why business needs it and who the competitor is

For AI to make real money in B2B, it has to log into a CRM on its own, analyze data, compile a report, and send it to a client without asking permission at every step. That is exactly the scenario Li Hang's framework addresses. ByteDance, which handles enormous volumes of data and runs complex recommendation algorithms, has more reason than almost anyone to want such autonomous systems. The work is framed as a direct challenge to Microsoft's AutoGen project. In the author's view, Western companies tend to rely on the "raw power" of compute, while ByteDance is betting on a structured, elegant, and scalable engineering design.

Frequently asked questions

What exactly did Li Hang propose?

A general-purpose framework for AI agents based on four modules: perception, planning, memory, and action. Instead of generating responses "in the moment," the model first builds a hierarchical tree of goals, then achieves them through external tools while continuously checking against long-term memory.

How is this different from AutoGPT and BabyAGI?

As the article describes them, existing agent systems such as AutoGPT and BabyAGI are unstable: they break in the early steps of a task and fall into hallucination loops. Li Hang's framework introduces a systematic architecture for planning and memory that is meant to turn these "toy" agent prototypes into a reliable industrial-grade tool.

Why could this be a challenge for Microsoft AutoGen?

Both projects tackle the same problem — building autonomous, multi-step AI agents — but, in the author's view, they take different routes. AutoGen and other Western efforts tend to lean on compute power, while ByteDance, through Li Hang, offers a structured engineering design. The outcome will depend on who moves the architecture from an academic paper into production faster.