The new product will compete directly with Anthropic's Claude Code and OpenAI's Codex.

According to Chen, the team will build Code Harness "from scratch." He described the project informally as "DeepSeek Code or something like that." The job listings use the formula "model + harness = agent" — a clear signal of the product's architectural direction.

The harness refers to a layer built on top of a large language model. It handles context management, step planning, memory, tool calls, file operations, terminal access, testing, and feedback loops. The goal is not a conventional code completion assistant but a full agentic product capable of executing multi-step tasks within a development environment.

Candidates are expected to have hands-on experience with Claude Code, Codex, Cursor, GitHub Copilot, Manus, and OpenClaw. Required skills include a solid understanding of agentic loops, Model Context Protocol (MCP), multi-agent systems, context management, and vibe coding. The team will work in close collaboration with DeepSeek's core researchers.

DeepSeek V4 already supports integration with Claude Code via an Anthropic-compatible API. The V4 Flash version, released on April 24, is priced at $0.14 per million input tokens — compared to $15 per million for Claude Opus 4.7, making the cost difference more than 100x.