DeepSeek Harness Is Open Source: Everything Is a Plugin

After DeepSeek V4 Pro's release, DeepSeek Harness (developer preview) went open source. The repo structure reveals it is not just another Codex, but a breadboard-like system. The system comprises various components like filesystem, terminal, subprocess, PTY, language servers, web access, skills, subagents, workflows, plan mode, session persistence, settings, credentials, and telemetry. Almost every capability has its own package, making it highly modular and customizable.

The project's name, "Harness," signifies its purpose: to connect the AI model to various mechanisms, such as the filesystem, shell, code editor, web, and other agents, while recording its activities and ensuring safety. The code volume justifies the existence of numerous packages, with only three key questions justifying their creation: can tool calls run in parallel, does cancel actually kill the subprocess, and do tool results pollute the context.

Harness operates on the Cordis microkernel, where each running Harness is essentially a Cordis Context. Packages register services, events, and capabilities into the context, and a config file composes them into a functional agent. The project consists of three layers: interface, implementation, and consumer. For example, Bash's interface defines what "execute a command" means, while the implementation spawns the process.

The model-facing tool package translates this capability into a schema and results understandable by the model. This separation allows for flexibility and ease of modification, as changing the implementation layer does not require rewriting the model tools or the agent loop.

The agent loop is not a traditional loop, but rather a traffic control system. Early agent projects typically involve sending messages to the model, executing tool calls, and sending results back, repeating until text. Harness, however, takes a stricter approach by defining a lifecycle for turns, steps, and steps, with tools passing through pre-policy, execution, post-processing, content normalization, and result notification stages. This allows for better control and safety in handling tool calls and their effects on the system.