Summary: DeepSeek shipped an MIT-licensed developer preview of DeepSeek Harness on August 13 as an open agent runtime built around a plugin system. It provides replayable append-only run logs and four modes aimed at daily coding, batch automation, benchmarks, and plugin development. The release lowers the cost of inspecting and customizing agent behavior, though the preview’s APIs can still change.
DeepSeek shipped a developer preview of its agent harness on August 13, calling it DeepSeek Harness, with the full source released under an MIT license. The project site leads with the thesis in plain terms: “Agent = Model + Harness.” A model is the soul of an agent, but the layer that lets an agent read its environment, use tools, and keep working in the real world is the harness. The release drew 747 points on Hacker News, putting it among the week’s most-watched developer tool launches. Setup is a single command — npx @deepseek-ai/dsh web (Node.js required) — and the code lives at the deepseek-ai/deepseek-harness repository on GitHub.
The Cordis Kernel: Everything Is a Plugin
The harness rests on the Cordis kernel, which takes care of mounting and unmounting plugins and of managing the dependencies between them. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI — every one of them is a plugin. Any of those capabilities can be picked, swapped, or extended from configuration, with no edits to the DeepSeek Harness source itself. In practice that means the differences between “our agent” and “the stock agent” can live entirely in a config file and a plugin directory: swap the model provider, replace the sandbox, add a scheduling loop, and your fork stays mergeable upstream. The design makes it feel less like a single CLI product and more like an assemblable operating system for agents. Cordis itself is a mature kernel with an existing community, which shortens the path from idea to working plugin.
Run Logs You Can Replay
Traceability is the second pillar. Everything the model sees — system prompts, reasoning, tool calls and results, subagent scheduling, context injections — lands in an append-only session log. Resume, fork, search, and replay all work off that one event stream. For debugging and evaluation this beats reading a transcript after the fact by a wide margin: a failed run can be replayed from any point with a different tool result or prompt, and a shift in behavior can be bisected across the event history. It also turns “why did the model do that” from an argument into a question you can replay step by step. It is the same idea version control applies to source code, applied to agent behavior: the run is a log of events, so any question about the run becomes a query over that log.
Four Runtime Modes, Four Jobs
The preview offers four modes. Standard is the complete coding agent: file editing, shell, search, skills, planning, goals, subagents, and workflows. Code hands tools to the model through a Code Mode SDK, which lets it chain multi-step operations into one TypeScript program that executes in a single pass — fewer round trips, and fewer openings for the model to lose the plot between calls. Minimal keeps only persistent bash and str_replace_editor, with the aim of benchmarking a model in a bare environment; it is effectively a reproducible harness for model evals. Creator brings runtime inspection, in-memory plugin testing, and preset authoring. Four modes map onto four jobs: daily development, batch automation, benchmarks, and plugin development. Because all four sit on the same plugin system, a skill learned in Minimal carries into Standard unchanged.
The Harness Race Has Started
The industry signal from this release is hard to miss: model companies now run the harness as a product line of its own. That layer used to be filled by third-party toolchains; now DeepSeek has stepped in directly, and from day one with an MIT license and a plugin ecosystem — community plugins are already gathering under the dsh-plugin tag, and the Cordis kernel has a corresponding paper. The company is upfront about the caveats: the preview is still under test, core plugins and APIs will keep changing, and teams that pin it for production should expect breakage. Even so, direction outweighs version number. If the model is the soul of the agent, the harness is the body — and until now, most teams building on frontier models had to accept a closed body they could not inspect. An MIT-licensed layer you can swap wholesale moves the default line: entry has never been cheaper, and lock-in has turned from a given into something a team chooses.