Kai-Fu Lee's "Super Employee": 01.AI's Wanzai Puts Agents at the Workstation

Over the National Day holiday, several Chinese model companies kept up the pace. Moonshot AI posted a registration entry for kimi-k3-1 in its API backend, StepFun released Step 5 Preview, and Kling opened version 4.0 for beta testing. 01.AI picked a different angle: rather than competing on parameter scale, it put its product straight into the office.

In early October, Kai-Fu Lee's 01.AI launched Wanzai, its first agent product, in Beijing, alongside version 2.0 of its Wanzhi enterprise large-model integrated platform. The company positions Wanzai as a "super employee." The phrase travels well at a launch event, but what it really means is something else: moving agents from executing a hard-coded workflow to drawing up their own execution plans.

What Wanzai actually does

According to what 01.AI has disclosed, Wanzai's capabilities fall into four buckets. It must do deep reasoning and planning for complex tasks, connect to phones and networked devices, plug into a company's existing service systems, and be tailored to specific business needs.

The first two determine whether it can "act"; the last two determine whether it can "land." The two questions enterprise buyers ask most often map exactly onto the latter half: where does the data live, and who verifies the results. Wanzai's answer is support for private deployment, plus a results-verification mechanism.

That sounds like table stakes, but in Chinese enterprise deals, private deployment is less a bonus than a threshold. Customers in finance, energy and manufacturing won't hand core processes to an assistant that can only run on someone else's cloud. The Wanzhi platform 2.0 serves as that foundation: models, knowledge bases, permissions and auditing all sit inside the customer's own environment.

The release rhythm itself is telling. 01.AI didn't put out a cloud trial first and then talk about an enterprise edition; it brought the private-deployment option out together with the product. That shows it isn't aiming at the consumer playbook of acquire first, convert later — from day one it has been building against an enterprise procurement checklist.

A question that keeps coming up

Over the past year, the complaint enterprises voice most about agents is that they "dare not let go." An agent that can call APIs and edit files on its own needs only slightly loose permissions to copy one mistaken action across ten systems. From late September into early October, almost every move at home and abroad revolved around this: some vendors are working on scope definitions, some on tamper-proof audit trails, some on millisecond-level circuit-breaker isolation.

By writing "results verification" into its product description, Wanzai concedes the same point: no matter how strong the capability, you must first answer "what happens when it goes wrong." For buyers, that carries more weight than a score on a leaderboard. A model ranked second that can trace and roll back every action will get into production more easily than the one ranked first.

From executor to decision-maker

At the launch, Lee drew a contrast. Common AI tools are built to run a predefined workflow to completion; Wanzai combines an enterprise knowledge base with production tasks to generate an execution plan on the fly. He described the shift as moving from a "tool-flow executor" to a "human-machine collaborative decision-maker."

The difference widens in practice. If an assistant only presses buttons, a human must first break the workflow down, and all it saves is clicking time. If it can decide for itself which systems to call, in what order, and at which point to hand back to a person, then what it saves is the mental effort of decomposing the process. That latter case is where the value of the phrase "super employee" lies, and it is the dividing line for whether such products are accepted at a real workstation.

On the technical side, 01.AI emphasizes in-house reinforcement learning and full-stack R&D to improve task planning. Claims like these can't be verified in isolation; what actually speaks is task completion rate and how often humans have to intervene — and those only count once they've been run on a customer's site.

Three rungs

Lee divides agent evolution into three stages. L1 is the workflow agent: humans assign the task, autonomy is limited. L2 is the reasoning agent, able to plan tasks on its own — the level he places Wanzai at. L3 is multi-agent collaboration, with multiple agents dividing labor across a decentralized network.

Placing the current product at L2 is a pragmatic call. In 2026, the vast majority of enterprise agents that actually work are stuck at this level: they can plan, but they need permission boundaries and a human safety net. What's genuinely scarce is who bears the cost when planning goes wrong. The launch didn't dwell on this, but it determines how long the road from L2 to L3 will be.

He also noted that Wanzai's capabilities are on the same order of magnitude as the ChatGPT Agent technology OpenAI previously announced. The analogy helps laypeople understand, and it also means 01.AI has placed itself within the same global comparison framework for agents.

A results-oriented signal

Put Wanzai back alongside the other moves over the National Day break and the signal is fairly clear. Domestic large models are shifting from talking about service to talking about results. Kimi K3.1 bets on long text, Step 5 Preview bets on efficiency, Kling 4.0 bets on video, and 01.AI bets on the process itself. Four companies, four paths, but pricing and deployment are tightening along the same line: whoever can save a customer one hire and one approval step is the one who gets the renewal.

"Super employee" is a good pitch and a high-bar promise. What enterprises are really buying is a piece of work that can be audited, held accountable, and replaced; a talking assistant is just the ticket in. Whether Wanzai can deliver depends on whether it dares to lay its execution records open for customers to see.