What was released

DeepSeek released DeepSeek Harness v0.1 in developer preview and published the full source code under the MIT licence. The project ships as dsh at deepseek-ai/deepseek-harness.

A harness is the layer between a model and the environment it acts in: the tools, files, sandboxes and control loop that let an agent keep working. DeepSeek frames it in a single equation: Agent = Model + Harness.

What makes it different

Most harnesses hard-code that layer. The agent loop, the tool registry and the session store are fixed, and extension happens only at whatever hooks the authors exposed. Harness takes the opposite position, and declares it in the first lines of its README: everything is a plugin.

The components placed behind plugin boundaries are:

  • models, tools and skills,
  • sessions, sandboxes and storage,
  • loops, scheduling and the user interface.

All of them sit behind Cordis plugin boundaries, and any can be selected, swapped or extended in configuration without changing Harness source code. That makes it a kit for assembling agent runtimes rather than a fixed coding assistant.

The Cordis kernel

Harness runs on a meta-framework called Cordis. The kernel's job is limited to mounting and unmounting plugins and handling dependencies. Capabilities live in the plugins, not in a privileged core.

That distinction matters architecturally. A kernel that holds no capabilities reduces the chance that changing one component breaks another, and it lets in-house tooling be packaged as reusable plugins.

Who it is for

The positioning is stated plainly: deployable, but as developer infrastructure rather than a production agent product. v0.1 is a developer preview.

The intended applications include internal coding agents over private repositories, model evaluation inside a controlled two-tool environment, agent observability and run replay, and custom sandbox and approval policies. Being MIT-licensed and self-hosted also lets regulated enterprises pilot it locally.

The real news here is the release itself rather than the model announcement it shipped alongside. The agent debate has long run on model capability; making the harness open source and fully swappable makes it easier to measure how differently the same model behaves across setups. It is worth keeping in mind that this is a developer preview and independent trials have not yet accumulated.

On the practical side, this design has a cost. Putting every component behind a plugin boundary makes setup more involved than a fixed assistant: the decisions about which plugin to select stay with you. In exchange, the same flexibility makes it possible to hold the harness constant and change only the model in comparisons, separating how much of an agent's performance comes from the model and how much from the layer around it — something that has been hard to measure until now. The repository being MIT-licensed also means that measurement can be done openly.