DeepSeek hires Beijing team to build Code Harness

DeepSeek posted job listings on May 20, 2026 for a product manager and an R&D engineer based in Beijing to build a product named Code Harness. The listings appeared on a company engineer’s post on X and on High-Flyer’s recruitment platform. Salary details were not disclosed.

The job documents describe an integrated “harness” that would sit on top of DeepSeek’s models and manage developer-facing workflows. The materials include the formula “Model + Harness = Agent.” The harness is described as providing command-line checkpoints, integrations and rollback features to support persistent, loop-heavy coding workflows.

The listings say the product would aim to autonomously plan, write, test and debug software projects rather than only suggesting individual lines of code. The product manager role requires hands-on experience with Claude Code, Codex, Cursor, GitHub Copilot, Manus and OpenClaw. Both roles are explicitly listed as Beijing-based rather than remote or based in other Chinese tech hubs.

Deli Chen, a DeepSeek engineer, shared the listings on X and wrote, “You can call it DeepSeek Code or something,” followed by laughing emojis in the post.

DeepSeek’s V4 model family is already integrated to run natively with Claude Code, according to company materials. The V4 lineup launched on April 24, 2026. V4 Flash is priced at $0.14 per million input tokens, and V4 Pro is listed at $0.435 per million input tokens under a promotional rate running through May 31, 2026. Anthropic’s Claude Opus 4.7 is listed at about $15 per million input tokens in the materials referenced by the listings.

The recruitment pages frame owning a harness as a way for DeepSeek to control the developer interface where usage and billing occur. No product launch date for Code Harness was included in the job postings or public materials.

DeepSeek was founded out of quantitative investment firm High-Flyer. The company released an R1 reasoning model in January 2025 and the V4 lineup in April 2026. The job documents and public posts indicate the company is building toward a full-stack offering that includes the model, an execution harness and a developer interface.