DeepSeek opens 150 engineering roles — and makes agent fluency mandatory
Two signals from China this morning: the open-weight lab is hiring like a platform company, and a phone is about to ship with an agent as its primary interface.
DeepSeek's Harness team is hiring roughly 150 engineers at once, and the job post makes fluency with AI agents a stated requirement rather than a nice-to-have. Cui Tianyi, who leads DeepSeek's Harness team, posted the call for senior backend and server-side engineers, saying the volume of data, machines, training jobs and evaluation jobs has grown past what the company's current backend systems can carry. The roles split into server-side development across six tracks — model research platform, agent framework components, developer productivity, the DeepSeek API, online serving and data engineering — plus agent elastic compute work on DSec, DeepSeek's in-house elastic compute platform for agents. Locations are Beijing and Hangzhou, targeting engineers with two to ten years of experience, though the team says fresh graduates are welcome too.
What makes this more than a headcount story is the requirement list. DeepSeek says candidates must be able to write production-quality code in languages, frameworks and domains they have no direct experience in, with agent assistance; they must own technical decisions in complex scenarios even when agents do the implementation; and they must be able to catch a bad agent proposal and intervene. That is a job description for supervising agents, not writing code around them, and it is the clearest signal yet that frontier labs now treat agent supervision as a core engineering competency rather than a productivity trick.
The context matters. DeepSeek is the open-weight champion that built its reputation on research efficiency, and it already said in June it planned to at least double the size of every department. This push is the infrastructure bill coming due: agents are compute-heavy and stateful in ways that chat serving never was, and the elastic compute platform that powers them is now a first-order product. A lab that once competed on how cheaply it could train is now competing on whether its systems stay up while millions of agent runs churn underneath them.
ZTE's Nubia brand will launch the NaviX Ultra on September 16, billing it as the world's first mass-produced "AI agent phone," with ByteDance's Doubao assistant built in. The pitch is that a single spoken request gets a task done end to end rather than opening an app, and Nubia has defined four capability dimensions it says define the category: response, execution, memory and security. The company also notes it completed a generative-model filing on July 15 and has since received its network access license, so the agent features ship with the regulatory paperwork China requires already in hand.
This is the consumer version of the same bet DeepSeek is making on infrastructure. Nubia shipped a technical preview of the Doubao-assisted M153 back in December 2025, and NaviX Ultra is the jump from demo to volume product. The interesting question is the security dimension: an agent that executes tasks on your behalf needs broad access to your apps and data by design, and a phone is the most permission-dense device most people own. Hardware, pricing and availability come at the September 16 event.
A Chinese man was admitted to intensive care after an AI vision tool told him a wild mushroom he had picked was safe to eat. He developed acute kidney failure and rhabdomyolysis after cooking it. Chinese media contrast the case with a Hong Kong woman who used AI to identify a rare toxic crab mixed into a plate of seafood and threw the whole dish out. Experts quoted in the coverage make the point plainly: image models match pixel statistics against training photos, and deadly and edible mushrooms can differ by a detail that lighting, angle or a smear of dirt erases — and no image model can detect a toxin it cannot see.
The failure mode here is not exotic hallucination, it is category error. A model that ranks photographs by similarity is being asked to answer a chemical question, and it will answer it confidently whether or not it knows. That is worth remembering on a day when two of the other stories are about handing agents more autonomy: the safe use of these systems still depends on someone knowing which questions they cannot answer.
What to watch: whether Nubia publishes a real permission model for the NaviX Ultra on September 16, and whether other labs follow DeepSeek in writing agent supervision into their hiring bar.
Have you ever trusted an AI answer on something that could actually hurt you — and did it change how you use these tools? Tell us in the comments.
Sources: 智东西 Zhidx · South China Morning Post · CNMO · 中华网 China.com · 杭州网辟谣平台
