DeepSeek's first agent product, DeepSeek Harness, is officially out — and the package is already live on npm. If you saw the public Code Agent evaluation of V4-Flash, that was Harness running its built-in minimal mode.
Harness isn't just another Codex-style coding agent. It handles project management, long-running tasks, multiple agents, Skills, web search, and context management, and it lets you create different "agent presets." Coding, research, and writing tasks can each get their own set of tools, prompts, and workflows.
The plugin system is where things get especially interesting. DeepSeek's core design principle is "Everything is a plugin": models, tools, prompts, storage, context, even the interface itself can be swapped or combined via plugins. Closed beta users have already built plugins for long-term memory and UI tweaks.
In short, Codex is a finished agent; DeepSeek Harness is more like a runtime for assembling one. You're not just choosing a model — you also decide which tools your agent uses, what rules it follows, and how it works.