DeepSeek has released DeepSeek Harness in developer preview, an open-source framework for building AI agents in which every capability — from the model itself to the user interface — is a plugin that can be swapped or recomposed. Source code is included, and the project is available on GitHub at `deepseek-ai/deepseek-harness`, with installation as simple as `npx @deepseek-ai/dsh web`.

The release, announced August 13, 2026, landed with immediate traction in the developer community: the launch page topped 700 points on Hacker News within a day, making it one of the most-discussed AI stories of the week. For ongoing coverage of the tools reshaping how developers build with AI, follow AI Buzz Wire.

Agent = Model + Harness

DeepSeek's framing for the project is disarmingly simple: the model is the soul of an agent, and the harness is what lets it work in the real world. A harness lets an agent understand its environment, use tools, and keep working in real-world settings rather than demo conditions.

The architectural centerpiece is the Cordis kernel, which manages plugin mounting, unmounting, and dependencies. Agent capabilities live in the plugins, and Cordis services and events let those plugins work together. The categories of capabilities exposed as plugins include:

  • Models — the reasoning engine itself is replaceable
  • Tools and skills — what the agent can do
  • Sessions and storage — how state persists
  • Sandboxes — where code and actions execute safely
  • Loops and scheduling — how work is organized over time
  • The UI — even the interface is composable

Developers can select, swap, or extend any capability through configuration alone, without changing the DeepSeek Harness source code. The approach is a deliberate bet that agent infrastructure should look less like a monolithic product and more like an ecosystem of interoperable parts.

Every Run Is Traceable

The second design pillar is transparency. Everything the model sees is recorded in an append-only session log: system prompts, reasoning, tool calls and results, subagent scheduling, and every context injection.

A built-in Trajectory view lets developers inspect these records by source, and crucially, resume, fork, search, and replay all operate on the same event stream. That means an agent session is not a black box — it is a durable, auditable record that can be examined and re-run after the fact. For teams debugging why an agent took an unexpected action, or auditors verifying what data influenced a decision, the feature addresses one of the most persistent pain points in agentic AI.

Four Runtime Modes

DeepSeek Harness ships with multiple runtime modes tailored to different needs:

Standard mode

A full coding agent with file editing, shell, file and web search, skills, planning, goals, subagents, and workflows.

Code mode

All Standard mode capabilities, with tools exposed through the Code Mode SDK so the model can combine multi-step operations into a single TypeScript program — letting the agent orchestrate multiple rounds of tool calls in generated code rather than one call at a time.

Minimal mode

A deliberately stripped-down two-tool coding agent with persistent bash and a `str_replace_editor`, intended for benchmarking models in a minimal environment.

Creator mode

Built for creating custom agent presets, with all Standard mode capabilities plus runtime inspection, in-memory plugin experiments, and preset-authoring guidance.

The four modes signal who DeepSeek expects to use the harness: developers shipping production agents, researchers benchmarking models under controlled conditions, and builders constructing entirely custom agent configurations.

Open Source as Strategy

The release continues DeepSeek's pattern of shipping substantial infrastructure into the open — a strategy that has repeatedly put the Chinese lab at the center of developer conversations since its model releases began drawing global attention. Alongside the harness, DeepSeek has published developer docs, a community plugin collection, and the Cordis research paper underpinning the kernel design.

By making the model itself just another plugin, the project also invites a provocative possibility: harnesses like this could become model-agnostic substrate, where OpenAI, Anthropic, Google, or open-weight models are interchangeable components in the same agent skeleton. Whether the community adopts Cordis as that standard — or treats it as one harness among many — will play out in the plugin ecosystem that grows around the GitHub repository in the coming months.

The developer preview is available now for agent harness developers worldwide, with source code included.

Stay Ahead of AI

For more reporting on the agent ecosystem and the tools developers actually use, bookmark AI Buzz Wire and check our latest AI tools coverage.

Read more AI news →