DeepSeek open sources architecture for proactive coding agent

A blueprint for proactive agents

DeepSeek has released a research prototype designed to show how a proactive coding agent can operate. The project, titled dsh-superagent-prototype, was published on October 31, 2026, as a reference for the research community. It is built on top of a separate framework called Harness, which the company released earlier in August.

The company emphasizes that this is not an official product intended for production use. DeepSeek notes that some ideas described in their accompanying design article are not actually implemented in the published code. It serves as a way for the public to see the architecture behind their approach to autonomous software tasks.

By releasing this prototype, DeepSeek is providing a starting point for others to see how agents can move beyond simple chat. The goal is to show how an agent can handle complex workflows rather than just answering isolated questions. This release signals a shift toward providing the infrastructure for agentic behavior.

The Harness and Cordis framework

The underlying framework, known as DeepSeek Harness or dsh, treats various components as composable plugins. These include model adapters, tools, sessions, sandboxes, storage, and the agent loops. This modularity allows developers to swap out different parts of the system depending on their specific needs.

The plugin runtime powering Harness is called Cordis. This framework is designed to dynamically compose different components and, importantly, reverse their effects when they are unloaded. This technical approach ensures that the system remains flexible while maintaining a structured way to manage various tools and environments.

Users can run a local web interface by installing a supported version of Node.js and using a specific command to launch the dsh web UI. This setup allows for a local environment where developers can inspect and modify how the agent interacts with different software components. The architecture is designed to be highly customizable for different types of tasks.

Router-and-Worker design

The Super-agent prototype introduces a specific design pattern called Router-and-Worker. In this setup, a Router evaluates incoming requests from platforms like GitHub or Feishu to determine if they are permitted. Once a request is approved, the system assigns it to an isolated Worker session to carry out the actual task.

This design moves away from an unrestricted autonomous loop, which can sometimes be unpredictable. Instead, it focuses on event-driven repository work where authorized requests or team-defined rules create specific work items. Each work item runs in its own dedicated agent session and directory before returning a result to the original conversation.

The system can potentially generate a pull request as a final output of these tasks. A web panel is also included to surface any work that requires human attention. This structure allows the agent to perform heavy lifting in the background while keeping a human in the loop for critical decisions or approvals.

Open source and security risks

DeepSeek released both the Harness framework and the Super-agent prototype under the permissive MIT license. This allows developers to clone, inspect, modify, and self-host the code freely. It also means developers can choose which underlying model they want to supply to the framework, separating the tooling from the specific AI model used.

However, the company includes several warnings regarding the current state of the software. They state that the prototype is experimental and has not undergone a security audit. They also note that the sandboxing and approval controls provided in the code do not guarantee total isolation from the host system.

Because of these risks, DeepSeek advises users to review all plugins and generated commands carefully. They recommend running the software with minimal privileges to protect the host machine. The release is intended for research and exploration rather than being an enterprise-grade solution ready for immediate deployment.

The competitive landscape

This release comes at a time when China's AI models are rapidly closing the gap with US frontier models. According to Epoch AI's Capabilities Index, models from DeepSeek, Qwen, and Kimi have significantly narrowed the performance gap seen in 2023. These benchmarks include complex tasks like FrontierMath and SWE-Bench Verified.

By open-sourcing the agent architecture, DeepSeek is positioning itself as a provider of the foundational plumbing for the next generation of AI tools. This creates a new type of competition for startups that build orchestration layers on top of commercial models. They now have a reference implementation from a major player to contend with.

The separation of the agent tooling from DeepSeek's model API business is a notable strategic move. It allows the company to foster an ecosystem where others can build on their architecture. While Anthropic has previously accused DeepSeek of distillation attacks, this open-source push focuses on the structural side of AI development.

Technical specifications

The prototype is hosted in a dedicated repository called dsh-superagent-prototype. This separation from the main Harness repository suggests that DeepSeek views the agent as a specific application of the broader framework. It allows the core framework to remain stable while the agent logic can iterate more quickly.

The architecture is built to handle multi-session interactions where workers return results to a primary thread. This mimics how humans might delegate tasks to assistants who work in separate environments. It ensures that the main conversation remains organized even as complex background tasks are being executed.

Developers interested in testing the system can look for the specific Node.js requirements mentioned in the documentation. By running the local web UI, they can see how the Router-and-Worker logic handles simulated requests. This provides a hands-on way to understand the mechanics of proactive agent design without needing to build the framework from scratch.