DeepSeek AI has released DeepSeek Harness v0.1 as a developer preview, publishing the full source under an MIT license at deepseek-ai/deepseek-harness. The project defines an agent as the combination of a model and a harness, where the harness mediates interaction with tools, files, sandboxes and the control loop. Unlike traditional harnesses that hard‑code the agent loop and tool registry, DeepSeek Harness treats every component as a plug‑in.
Built on the Cordis meta‑framework, the harness kernel manages plugin mounting, unmounting and dependencies while all capabilities—models, tools, skills, sessions, sandboxes, storage, loops, scheduling and the UI—reside in plugins that can be selected, swapped or extended via configuration without touching core code. Four runtime profiles are provided: Standard (full coding agent), Code Mode (SDK‑exposed toolset), Minimal (bash and editor for bare‑metal benchmarking) and Creator (runtime inspection and plugin authoring). All interactions are appended to an immutable session log enabling replay, fork and search.
- MIT license, v0.1 developer preview
- Cordis plugin architecture
- Profiles: Standard, Code, Minimal, Creator
- Session log records prompts, reasoning, tool calls, context injections
- UI:
npx @deepseek-ai/dsh web(http://127.0.0.1:3080) - SDK: Python 3.10+, Linux/arm64/macOS arm64, no Node required
Why this matters
The plugin‑first architecture decouples model inference from orchestration logic, allowing researchers to exchange LLMs or tools without recompiling the harness. This modularity speeds up experimentation with diverse models and custom tooling while the exhaustive, append‑only session log guarantees reproducible traces for debugging and audit.