2B agents are not copies of 2C: enterprises will let AI hack their systems
The Afine founder believes 2C personal agents are oversaturated, competing on supply chains and subsidies; 2B agents more easily capture end-to-end value because bosses actively cooperate with AI to transform their companies.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
A hundred agents plus a hundred employees amplify noise
When you give each employee a Personal Agent, leadership no longer faces 100 employees but 100 employees plus 100 uncoordinated agents, amplifying noise rather than signal. Their solution is a single agent shared across the company with continuous learning that preserves individual memory and operating environments while learning across all business touchpoints.
— He Jiachen2C agents have almost no economies of scale
AI startups differ from internet startups because scale and network effects are greatly weakened—more users doesn't make your model smarter; the model's own intelligence improvement is decisive. This is why they don't build 2C personal agents: that market is about supply chains and subsidies, not product moats.
— He Jiachen2B bosses actively let AI hack their own systems
In C2C, getting all user information for recommendations is frightening; agents naturally need tricks and reverse engineering to get permissions. In B2B, when the boss sees that AI can trace end-to-end business data, financial data, and every code flow, the boss has strong incentive to actively transform systems to give AI more context. This makes 2B agents easier to adopt early on.
— He JiachenWorkflow plugins have shorter and shorter lifespans
Products like LongChain, Dify, N8N, and Open Cloud that scaffolded workflows once soared in valuation but are barely mentioned by H2 2026. Not because better workflows replaced them, but because models absorbed these workflows—with smarter models, the first move is usually to delete previous skills and system prompts or they backfire. Determinism is shifting from outside the model to inside, through KV Cache and Prompt Cache hit rates.
— He JiachenBuilding your own models only became viable in 2026
The speaker deeply participated in training a domestic ET-scale multimodal agentic model and didn't believe startups could train models—big tech has vastly superior chips, talent density, and end-to-end infrastructure. But by 2026, with architecture changes (like Jamba models) and base model capabilities reaching threshold, small models don't need to beat AGI models; they just need to efficiently distill tasks they can definitely do at lower cost and faster latency. That makes startup model training viable.
— He JiachenSmall models are the autonomic nervous system, large models are the cortex
They train two types of models. Task Routing Models distill the tool-calling patterns and user preferences from SOTA models into ‘muscle memory,’ letting small models form conditioned reflexes like learned bike riding. Worker Models of 20-40 billion parameters handle reverse-engineering and specific tasks at the operating-system level of enterprises, breaking through safety constraints. The intelligent large model is the cortex; their trained small models are the brainstem and spinal cord.
— He JiachenAll software today is essentially open source
If anyone knows what your software does and can get the API spec, AI can reverse-engineer it or rewrite an identical replacement. So software itself is dying, and value shifts to model intelligence itself. The speaker predicts that in 2-3 years, open models will catch up much faster, AGI supply will become abundant, and the commercial rules for both B2B and C2C will change very bloodily.
— He JiachenAgents are crossing the inflection point that video once owned
In the past 2-3 years, only chat and video—with predictable delivery and commercial closure—made money early. General agents need end-to-end completion to create value; one missing step means zero value. But that gap recently closed, and the advantages video and companionship tracks accumulated face challenge from general infrastructure and general models. OPUS 5.5 is the best video-editing model, GPT6 Astra the best at robotic video control—a complete reversal.
— He JiachenIn their own words · checked verbatim
This is also the core reason we didn't choose to build 2C personal agents.
这也是我们没有选择做2C的personal agent的核心原因
He Jiachen16:10
But in 2B scenarios, the enterprise's number-one executive is willing to let you hack their system.
但在2B的场景里面 企业的一号位负责人 就他是乐意让你去hack他的
He Jiachen19:13
Actually, these major clients have stronger motivation than me to make necessary changes to their systems so AI can get more context and do more work.
实际上这些大客户有比我更强的动力 去对他们的系统做必要的改造 就是为了让AI能拿到更多的context 进行更多的铲除
He Jiachen20:13
When you've been deeply involved in model training, during that time I was especially skeptical of the startup model training narrative.
当你深度的参与过模型训练之后 其实在那段时间 我是特别不相信创业公司的模型训练这个叙事的
He Jiachen35:31
We have two fundamental judgments; the first is, we think all software today is open source.
我们有两个比较根本的判断 第一是我们觉得说 今天所有的软件都是开源软件
He Jiachen40:34
OPUS 5.5 is the best video-editing model, and GPT6 Astra is the best at controlling robotic video—it's a complete reversal.
就OPUS 5.5是最会剪视频的模型 然后GPT6 Astra是最会控制Robotic的视频模型 这简直就是倒反天刚
He Jiachen43:34
Glossary
- Proactive Agent
- An agent that not only passively responds to conversation but proactively contacts users and advances tasks.
- KV Cache / Prompt Cache
- Storage of the model's previously generated results, enabling identical calls to be reused directly without re-inference.
- Task Routing Model
- A small model that distills the large model's task-completion trajectories and user preferences, replicating them at low cost.
How to listen
Founders focused on agent product evolution, decision-makers building 2B enterprise software or AI transformation, and practitioners curious how model training lands in startups.
The opening pleasantries about the founder's personal experiences can be skipped.