Skip to content

Open models

DeepSeek open-sources its agent harness under MIT

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: Coding agents →
Image: github.com

On August 13, DeepSeek open sourced DeepSeek Harness, a Node.js agent runtime in which every part can be replaced, including the model adapter, the tool registry, the session log and the agent loop. It runs on Cordis, a composition framework described in a paper by researchers from DeepSeek and Peking University.

The project is on GitHub as a developer preview under the MIT license, The New Stack reported. It passed 33,000 stars within a few hours, and community plugins were already appearing. It shipped a day after DeepSeek’s V4-Pro flagship went live, but nothing in it depends on DeepSeek’s own models.

No privileged core

DeepSeek’s phrase for the design is “everything is a plugin,” and it means that literally. The documentation says there is “no privileged core to patch.” To extend the harness, you mount a plugin next to the others and leave the central runtime alone. Swapping out the agent loop works the same way as adding a tool.

Cordis is what makes that safe. DeepSeek calls it a “meta-framework for spatiotemporal composability.” The paper behind it, “A Programming Paradigm for Spatiotemporal Composability,” is by Yifan Shi (Peking University and DeepSeek-AI), Wei Zhang (Peking University) and Tianyi Cui (DeepSeek-AI). It splits the problem in two. Temporal composability is “the ability to completely revert a component’s side effects upon removal.” Spatial composability is “the ability to declare and reactively manage inter-component dependencies.”

In code, it works like this. When a plugin changes the shared context, ctx.effect registers the change and records how to undo it. When the plugin is disposed, the runtime runs those undo steps, so unloading leaves nothing behind.

Dependencies are handled separately. A plugin lists what it needs in an inject declaration. Cordis works out the load order from those declarations and updates a plugin when something it depends on changes. The paper calls the two halves revertible effects and reactive coeffects. A coeffect is what a component needs from its context, and an effect is what a component does to it. Every context change is checked against each component’s coeffect specification, and the result decides whether the component switches on or off.

The framework is made of a core library that tracks effects and resolves dependencies, plus a declarative component loader with configuration reconciliation and hot module replacement. The paper’s case study is Koishi, an open-source chatbot framework built on Cordis that has been in production use for four years. The paper also lays out a formal calculus covering components, fibers, withdrawal, iteration and failure, with proofs of preservation, progress and confluence that extend the composability guarantees from one component to a whole system. For its first two weeks, the only public copy was in the cordiverse/paper repository, under the GitHub organization that maintains Cordis. Later, on August 26, the paper was submitted to arXiv under cs.PL and cs.SE.

Four presets

The harness ships with four presets. Standard is the full coding agent, with filesystem tools, shell access, web search, subagents and a plan mode. Minimal keeps only two tools, bash and str_replace_editor.

Code mode changes how tools reach the model. Tools aren’t exposed as separate function calls. The harness generates a TypeScript SDK and the model writes a program against it, so a sequence that would take five round trips runs as a single call.

Creator is for developers building their own presets. It inherits everything in Standard and adds runtime inspection, plugin experiments and guidance on writing presets.

One log behind everything

The harness keeps an append-only session log, and anything that reaches a model request must be reconstructable from it. Resume, fork, replay, transcripts, telemetry and the web UI all read from that one event stream. That makes conversation history more than an implementation detail. If you want the model to see a new kind of input, you first add a new session event type.

Sandboxing differs by platform. On Linux, the local backend wraps subprocesses in Landlock, the kernel’s sandboxing interface for unprivileged processes, through a Node addon DeepSeek wrote. On macOS it uses Seatbelt, and on Windows a runner built on ACL restricted tokens.

Claude Code and Codex as subagents

The provider catalog covers Anthropic, OpenAI, AWS Bedrock, Microsoft Azure and Google’s Gemini Enterprise Agent Platform, alongside DeepSeek’s own endpoint. The documentation still calls the Google platform Vertex. You can also add custom OpenAI-compatible gateways for other inference providers.

Two subagent providers hand work directly to Anthropic’s Claude Code and OpenAI’s Codex. Each finds the product’s binary on the host PATH, so you supply the install and the login yourself. Both are switched off by default. DeepSeek also ships bridges that run an existing hooks.json from either product against the harness’s own interception points, and the README calls them a compatibility path and says they are not the better design. The harness also has an MCP client and Agent Client Protocol support, and it reads AGENTS.md and CLAUDE.md files.

The field is crowded. Among Chinese labs, Alibaba’s Qwen Code and ByteDance’s Trae Agent launched in mid-2025, Moonshot’s Kimi CLI followed that October, and Zhipu’s ZCode arrived in July. The New Stack notes that most of the competition is open source as well.

Who gets to contribute?

The repository went past 33,000 stars in a few hours, yet none of those developers can get code merged into it so far. “We are sorry that we cannot accept external pull requests at the moment,” DeepSeek writes. It points would-be contributors to GitHub Discussions and suggests they build plugins.

DeepSeek says the packages in its official repository are not inherently more important than community ones. It asks readers to treat the repo as “an idea, an official showcase, and a source of inspiration, but not a mandate from us.”

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.