A phone on a Shanghai desk unlocks itself, opens a food app, reorders the usual, and pays — without a finger touching the screen. The assistant did not write a poem about dinner. It went and got it. That is the line Chinese agents crossed in 2025: from answering to acting.
The shift from chatbot to do-bot is where the large model (大模型) story stops being academic and starts rearranging industries and, quietly, geopolitics.
The agent that became a geopolitical object
Manus was the clearest signal. The Chinese-founded agent startup drew global attention with a viral demo of an AI that completed real workflows — research, filing, booking — rather than describing them. In December 2025 it agreed to a roughly US$2 billion acquisition by Meta. Then Beijing stepped in. Regulators forced the deal's unwinding on national-security and technology-export-control grounds, and Meta cut operational ties.
That episode is instructive for two reasons. First, it showed how far a Chinese agent product had travelled commercially in a year. Second, it revealed that Beijing now classifies capable agent technology as something that should not simply flow outward with foreign capital — the same instinct behind US chip controls, pointed the other way.
Frameworks that turn models into workers
The agents that matter most are not consumer novelties; they are frameworks that let a model plan, call tools, and recover from errors. Moonshot AI's Kimi K2.5, released in January 2026, arrived with Kimi Code, an open coding agent that accepts screenshots and video and competes with tools like Claude Code and Gemini CLI. Zhipu's GLM line pushed toward the same goal: its 2026 "Ox Alpha" release was positioned as a reasoning model built for coding and sustained agentic work, the kind of long-horizon software engineering that used to need a human.
Zhipu has also pushed its AutoGLM line — agents the company markets for controlling phones and desktop software rather than merely chatting. The bet is that the valuable layer is not the model that talks, but the agent that operates the apps people already use.
The on-device phone agent
The most visible version of "acting" lives in the phone itself. Chinese device makers are racing to put agents locally on hardware, where they can read the screen, understand intent, and act without round-trips to a cloud. Honor's MagicOS 9, for instance, ships an on-device call-translation feature that downloads a translation model straight to the phone and processes voice locally across six languages — speech that never leaves the device. It is a small agentic capability, but it shows the direction: inference moving onto the silicon in your pocket.
The strategic appeal is obvious. On-device agents dodge latency, privacy, and cross-border-data concerns all at once, and they turn a handset into a platform the manufacturer controls. For vendors locked out of the most advanced foreign chips, running capable agents locally is also a way to keep the experience competitive without scaling the cloud.
Why China moved faster than the West here
Western agents tend to live inside apps and APIs, gated by enterprise procurement. Chinese agents leapt straight to the device and the super-app, where a single assistant can reach payments, commerce, and messaging in one move. That is partly structural — China's app ecosystem concentrates functions that the West splits across vendors — and partly cultural, a tolerance for the assistant doing things on your behalf.
It is also a funding story. Zhipu, one of the anchor agents-and-models labs, had raised 2.5 billion yuan (≈ US$340M / HK$2.65B) in financing by late 2025 and later listed in Hong Kong with a market capitalisation around HK$434.7 billion (≈ US$55.9B); peer MiniMax reached roughly HK$257.3 billion (≈ US$33B) after its own listing. Capital is flowing to labs that can pair a model with an agent that uses it.
The risks that travel with autonomy
An agent that can pay for dinner can also make a costly mistake, expose a credential, or act on a confused instruction. The same autonomy that makes on-device agents useful makes them a liability surface. Honor's local-translation example is low-risk precisely because its action is narrow; broader agents that browse, book, and buy need guardrails the industry has not standardised.
There is also a trust question. Agents that operate apps on a user's behalf see everything on the screen. On-device processing helps, but the business models behind many agents still pull activity to the cloud.
Honest limitations
This article draws on tech-media reporting (TechCrunch) for Manus, Kimi Code, and Honor's on-device features, and on company disclosures for Zhipu's agent positioning and financing; the HK$ market-cap figures are as reported at listing and move with the market. The Manus divestiture and export-control framing reflect reporting at the time of writing and may evolve as the unwinding completes. Coverage focuses on the chat-to-action shift, on-device phone agents, and the strategic framing; it does not assess specific safety benchmarks for autonomous action, nor does it deeply compare Chinese and Western agent frameworks beyond the structural contrast noted. Claims about Zhipu's AutoGLM capabilities are the company's own positioning rather than independently verified feature tests.
What readers can do now
- Separate narrow agents from open-ended ones. Start with a constrained, on-device agent (like local translation or a single-app task) where the blast radius is small before trusting a broad agent with payments and accounts.
- Watch the Hong Kong listings as a signal. Zhipu and MiniMax valuations (HK$434.7B and HK$257.3B at listing) are a real-time read on how seriously capital treats the model-plus-agent bundle — track them alongside product releases.
- Prototype an agent on open frameworks first. Try Kimi Code or Zhipu's open agent tooling on a sandboxed task before granting any agent access to real accounts, credentials, or money.