DeepSeek doubles down on “Code Harness” as AI coding agents become the front line

Pivot to agents: from base models to Code Harness

It has been reported that DeepSeek — the AI unit associated with Chinese parent company 幻方 — is shifting its strategy from base-model glory to shipping agent-first coding products. After news of a reported RMB 70 billion financing round, the company has posted multiple job openings for an "Agent Harness" product manager and engineers. The listings make plain what the firm is building: a desktop Code Agent and the surrounding harness that turns model capabilities into reliable, long-running developer workflows.

Hiring shows product-and-systems emphasis

DeepSeek has reportedly recruited ex-Jane Street engineer Cui Tianyi (崔添翼) into a newly formed AI Harness team. Cui’s background in production trading systems — rather than academic model papers — signals DeepSeek’s need for people who understand how model outputs behave in real systems. At the same time the company has been hiring model-structure talent such as former ByteDance (字节跳动) researchers to shore up long-context and state-tracking capabilities. In short: one track builds execution and product robustness; the other fixes the model’s ability to preserve and update state across many steps.

A global arms race with commercial stakes — and geopolitical overtones

The move comes as Claude Code, OpenAI’s Codex desktop app and Google’s Antigravity 2.0 push coding agents into mainstream developer tooling. It has been reported that Anthropic’s Claude Code generates more than $2.5 billion in annualized revenue and that OpenAI’s Codex saw rapid weekly-active-user growth after its desktop launch. Against a backdrop of US export controls and broader technology decoupling, Chinese AI groups are racing to deliver homegrown agent ecosystems that do not rely on foreign cloud or tooling. Can DeepSeek turn heavy hiring and a Code Harness playbook into a commercial product that rivals Claude and Codex? The next year will answer that question.