DeepSeek Open-Sources Its Agent Harness
DeepSeek Harness packages a coding-agent runtime, Web UI, tools, subagents and SDK behind a plugin-first architecture, but its developer-preview label means adopters should expect breaking changes.
The 60-second version
DeepSeek Harness is an MIT-licensed, plugin-first agent runtime with Web, headless and Python interfaces.
Key points
- Model adapters, tools, sessions, prompt assembly and even the agent loop are plugins built on Cordis.
- The Web UI supports workspace editing, command execution, delegation, planning and approval-controlled actions.
- A headless profile handles one-shot tasks, while the Python SDK embeds a bundled runtime.
- DeepSeek and other catalog or custom OpenAI-compatible model providers can be configured.
- The project is a developer preview and explicitly makes no current compatibility promise.
Verdict. It is a substantial open agent platform for experimentation, but version pinning and isolated workspaces are mandatory until its interfaces stabilize.
DeepSeek has open-sourced the runtime around an AI agent, not another model. DeepSeek Harness, or `dsh`, combines a Web interface, model adapters, tools, durable sessions, approvals, subagents and SDK access. Its main architectural bet is that every major part — including the agent loop — is replaceable as a plugin.
Core ideaEverything is a plugin
Harness is powered by the Cordis plugin framework. Model providers, prompt assembly, tool execution, session logs and the default loop register services and events in a shared context. Profiles then compose ordered bundles with user overrides, allowing a developer to swap one capability without rewriting the rest of the runtime.
| Model layer | DeepSeek, catalog providers such as Anthropic and OpenAI, native cloud routes, and custom OpenAI-compatible endpoints. |
|---|---|
| Agent layer | Prompt assembly, the model request loop, tool calls, approvals, planning and delegation. |
| State layer | Append-only session events support persistence, replay, resume and derived transcripts. |
| Extension layer | Plugins can add providers, tools, commands, background work, sandbox policy and UI elements. |
Three entrancesBrowser, headless and Python
The shortest documented command is `npx @deepseek-ai/dsh web`, which starts a local Web UI at `127.0.0.1:3080` by default. Users configure a model, choose a workspace and create a session. The same base can also run a one-shot headless task with no server, or be driven programmatically through the Python SDK.
What it can doA full coding-agent surface
The Web guide says the agent can read and edit files in a selected workspace, run commands, delegate work and maintain a plan. The base bundle also includes persistence, credentials, telemetry, filesystem and shell policy, jobs, skills, Web tools and subagent providers. Which capabilities reach a session depends on its profile and plugin composition.
- Web UI: an interactive workspace, session history, model settings and approval prompts.
- Headless profile: one fresh persisted session for a task, with the final response written to standard output.
- Python SDK: a context-managed runtime with reusable session IDs and a checked-in JSON-RPC example.
- Custom composition: inspect the active plugin tree and override rows through profile patches.
Operational limitsPreview software needs containment
The repository's source toolchain targets recent Node.js and pnpm versions, while the Python SDK supports Python 3.10 or newer on documented Linux and Apple Silicon platforms. More importantly, tools may modify files and run shell commands. The Python guide's minimal composition uses `danger-full-access` and recommends a disposable checkout or container.
AdoptionWho should try it now
Harness is most compelling for developers studying agent architecture, building custom tools or providers, or comparing Web, headless and embedded execution over one runtime. Teams that need long-lived session compatibility or a frozen plugin API should wait, or isolate the preview behind their own adapter.
The release matters because DeepSeek has published an agent operating system, not merely a chat wrapper.
The practical move is to test it in a disposable project, pin the exact release candidate, inspect permissions and composed plugins, and treat every upgrade as a migration. The architecture is ambitious and unusually transparent; the compatibility promise is not ready yet.