On March 16 Alibaba carried out a near–foundational reorganisation and unveiled a bet that reframes how work will be automated inside Chinese companies. The group created ATH (Alibaba Token Hub) and promoted a previously unseen Wukong business unit into the company’s core sequence, with DingTalk’s chief executive presenting the new agent platform that Alibaba intends to bind into enterprises rather than leave as a consumer novelty.
The pitch is simple and sweeping: the operating主体 of many routine workplace tasks is shifting from people to AI. DingTalk’s CEO divided the history of human–computer interaction into four eras and declared that Alibaba is preparing for the next one by rewriting DingTalk’s base code so that its entire product capability is callable as atomic command‑line instructions. Messaging, calendars, approvals, CRM access, documents, meetings and hardware are being decomposed into thousands of CLI commands so an agent can invoke services directly rather than simulate user clicks.
Alibaba showed ten filmed case studies in which the agent, branded Wukong, completes entire job flows end to end. The examples ranged from running targeted local marketing for an auto repair shop to producing audited cash reports from 15 banks, generating working software from meeting recordings, and preparing pretrial legal briefs. The company calls the concept OPT (One Person Team): one human plus one Wukong can replicate a large fraction of a small team’s operational output.
What Alibaba emphasised as the crucial difference from the proliferation of hobbyist agents—nicknamed “lobsters” in domestic tech discourse—is institutionalisation. Wukong is designed to be “kept in a cage”: bound to a user’s enterprise DingTalk identity, constrained by multilayered safety controls, and auditable at scale. That contrasts with standalone agents running on personal machines where permission boundaries, data leakage and accounting for compute are unresolved.
The technical stack undergirding this vision spans identity, file handling, ecosystem integration and hardware. DNA authentication and security sandboxes attest agent identity and limit what commands and data an agent may access; enterprise‑only VPC deployments and multi‑stage runtime protections provide further containment. RealDoc is an AI‑native file system that enables fine‑grained edits and versioning to avoid the massive token costs of rewriting entire documents. An AI skills marketplace will let domain experts package capabilities for reuse and monetisation.
Alibaba also unveiled Realbox, a purpose‑built AI compute appliance, plus two personal AI devices: an A1 Pro recording card and Cleer H1 AI earphones. Realbox is pitched as infrastructure: a single unit hosts multiple PC and mobile runtime environments so businesses can run many Wukong instances in parallel and account for compute centrally. The strategy is to make compute, access and skills into enterprise utilities rather than ad‑hoc desktop experiments.
Strategically this is ATH as more than a product: a hub for token flows. Alibaba positions itself to mediate the generation and consumption of model tokens across its cloud, marketplace and commerce assets. By packaging skills and exposing underlying platform capabilities (Taobao, Tmall, 1688, Alipay, Alibaba Cloud) as CLI endpoints, Wukong aims to be the layer that electrifies traditional business processes with AI—while keeping the billing, security and governance under a single vendor’s umbrella.
The challenges are material. A polished demo across ten industries is not the same as convincing thousands of firms to rewrite workflows and trust mission‑critical data to new runtime paths. The depth of skills in the marketplace, Realbox’s real‑world throughput, cross‑platform reliability and the behavioural and regulatory shifts companies must make will determine if Wukong is a strategic platform or a sophisticated feature. For now, Alibaba has put its chips on enterprise containment, billing and integration rather than on merely shipping a more capable chatbot.