DeepSeek has released DeepSeek Harness in developer preview, making the source code available for developers building AI agent systems, according to DeepSeek materials surfaced on Hacker News. The project, also referred to as `dsh`, is an open-source agent harness: software meant to help an AI agent understand its environment, use tools, maintain sessions, and keep working across real-world tasks. The release is not being presented as stable. DeepSeek’s repository notes that Harness is in developer preview, is iterating quickly, and may introduce compatibility-breaking changes. That matters for teams considering it as infrastructure rather than a demo: the current version is suitable for evaluation and experimentation, but DeepSeek is explicitly warning that core interfaces may change. The central design choice is modularity. DeepSeek says Harness is built on Cordis, a plugin system whose kernel manages plugin mounting, unmounting, and dependencies. In the Harness model, agent capabilities live in plugins: models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the user interface can be selected, swapped, or extended through configuration without changing the Harness source code. That architecture is the clearest reason developers will compare Harness with other agent-building and coding-agent environments. VentureBeat frames the launch as an open-source rival to Claude Code, while DeepSeek’s own materials describe a lower-level harness for composing agent capabilities. The provided sources do not establish feature-by-feature parity with Claude Code, so the safest read is narrower: DeepSeek is moving beyond model releases into the software layer developers use to run and inspect agents. DeepSeek is also emphasizing traceability. Its preview materials say Harness records what the model sees in an append-only session log, including system prompts, reasoning, tool calls and results, subagent scheduling, and context injections. A Trajectory view lets developers inspect those records by source, and functions such as resume, fork, search, and replay operate on the same event stream, according to the DeepSeek page. Harness ships with several modes. Standard mode includes the full toolset. Code mode uses model-generated code to coordinate multiple rounds of tool calls. Minimal mode keeps only a shell tool and file editor for benchmarking models in a smaller environment. Creator mode is aimed at developers testing Cordis plugins in memory and combining them into new modes. The project can be started with Node.js and launched through `npx`, according to DeepSeek’s preview materials, with a Web UI served locally by default. Developers who want the full setup are directed to clone the repository and follow the project’s development and architecture documentation. VentureBeat reports that DeepSeek paired the Harness launch with the official version of DeepSeek-V4-Pro, described in its summary as an updated flagship model focused heavily on agentic workloads. VentureBeat’s headline also says V4-Pro is available through the API with higher prices, but the provided summary does not include the new pricing levels. That leaves the model and pricing side of the story single-source in this cluster, while the Harness preview itself is supported by DeepSeek’s own public materials. Who benefits: Agent developers and engineering teams evaluating open-source harnesses get another framework to test, particularly if they want configurable plugins and replayable session traces. DeepSeek also benefits if Harness increases adoption of its agentic model stack. Who's exposed: Teams building on more mature coding-agent workflows may face comparison pressure if Harness gains developer mindshare. Early adopters of Harness are exposed to compatibility-breaking changes while the preview evolves.