DeepSeek Harness

Quick Take

DeepSeek Harness (dsh) is DeepSeek AI’s open-source, MIT-licensed agent framework, launched as a developer preview in August 2026. Its core idea is that everything an agent needs, the model it calls, the tools it can use, the sandbox it runs in, its session logs, and the loop that drives it, is a plugin. Nothing is hard-wired, including the model: it’s genuinely model-agnostic, and can even call other coding agents like Claude Code or Codex as sub-agents inside its own orchestrated workflow.

Everything Is a Plugin

The framework’s architecture, described in DeepSeek’s own documentation as built on a “Cordis” design for spatiotemporal composability, treats every functional piece of an agent as a swappable component:

• Model adapters: the underlying model isn’t fixed to DeepSeek’s own; you plug in whichever model you want
• Tool registry: what tools the agent can call
• Session logs: how state and history are recorded
• Sandboxes: the execution environment an agent’s actions run inside
• Storage: where persistent data lives
• Loops and scheduling: how the agent iterates, retries, and decides when to stop
• The UI itself: both a browser-based web UI and a headless mode are supported

Because the model adapter is just another plugin, DeepSeek Harness isn’t locked to DeepSeek’s own models. It can call other coding agents, including Claude Code or Codex, as sub-agents inside a workflow it orchestrates, which positions it less as a rival to those tools and more as a layer that can sit above and coordinate them.

Why This Matters: The Harness Is the System Around the Model

A model on its own doesn’t make an agent. What actually determines how an agent behaves in practice is the harness around it: how prompts get assembled and how state persists between steps, what tools it can reach, what sandbox constrains it, how retries and verification work, and how the loop decides it’s actually done. DeepSeek Harness makes that usually-invisible layer explicit and, deliberately, entirely replaceable piece by piece. That’s a useful concept to understand even if you never run this specific project: see harness engineering for the general idea it’s built around.

Key Facts

• Developer: DeepSeek AI
• Status: Open-source developer preview, launched August 2026
• License: MIT
• Interfaces: Browser-based web UI and a headless (CLI-driven) mode
• Model support: Model-agnostic by design, including the ability to call other coding agents as sub-agents
• Community traction: Reported at over 160,000 GitHub stars and 18,000 forks shortly after launch, worth verifying against the live repository for current figures

Honest Limitations

• Developer preview, not a finished product. Expect breaking changes, incomplete documentation in places, and rapid iteration as it matures.
• A framework, not a turnkey agent. You still need to configure models, tools, and sandboxes yourself; it’s infrastructure for building an agent system, not a ready-made assistant.
• Requires developer comfort. This is aimed at people building agent infrastructure, not end users looking for a packaged tool.
• Track record is short. As with any newly launched open-source project, real-world reliability at scale is still being established; verify current status and activity on the GitHub repository before depending on it for production work.

Alternatives Worth Knowing

• Claude Code, Codex, and OpenCode, coding agents DeepSeek Harness can itself call as sub-agents rather than only compete with
• LangGraph and CrewAI, other frameworks for building and orchestrating multi-step or multi-agent workflows
• DeepSeek, the model family from the same lab, usable inside the harness but not required by it

For the underlying concepts, see Harness Engineering, Loop Engineering, Agent Sandbox, and Model Routing.