Deepseek releases harness: open-source agent framework where everything is a plugin, including the model

DeepSeek has made the first development version of DeepSeek Harness publicly available. The open-source framework, distributed under the MIT license and installed via the command dsh, is positioned not as a standard code-assistant but as a composable toolkit for developers. According to the documentation, an “agent” is equivalent to a model plus a harness, the intermediary layer between the model and its execution environment, and every component in this layer is intended to be replaceable.

The harness is built on Cordis, a meta-framework described in the paper “A Programming Paradigm for Spatiotemporal Composability”. Cordis’s core is limited to anchoring plugins and handling dependencies, while capabilities such as models, tools, sessions, sandboxes, storage and the user interface are all implemented as plugins. Consequently, developers can swap the model provider, the scheduling mechanism or the sandbox policy through a configuration file without modifying the system’s source code.

Four distinct runtime modes are defined. The Standard mode runs the full coding agent with file editing, shell control, search and sub-agents. The Code mode lets a model compose multi-step operations within a single TypeScript program via a dedicated SDK. The Minimal mode reduces the environment to only a bash shell and the str_replace_editor, enabling benchmarking of models in an exposed setting. The Creator mode adds in-memory plugin testing and the ability to write predefined configurations.

A highlighted feature is comprehensive logging. Every piece of information the model receives—including system prompts, tool calls, their results and any context injection—is written to an appended session log. The Trajectory view allows inspection of these records by source, and actions such as continuation, branching, search and replay operate on the same event stream. While most agent frameworks log tool calls, logging of full context injection is a rarer capability that is especially relevant for oversight and audit environments.

Model routing is also handled as a plugin. Supplying an API key for DeepSeek or providers such as Anthropic and OpenAI becomes effective on the next request without restarting the server. Providers like Bedrock, Vertex, Azure and Codex require native authentication, and custom providers can use an OpenAI-compatible base URL. Keys are stored in a separate, write-only file, so system settings retain only the identifier, not the secret value. The release targets development infrastructure rather than a finished product, with the primary audience being startups and internal development teams that already manage proprietary tooling.