DeepSeek Harness: What "Everything is a Plugin" Actually Means for Agent Frameworks

DeepSeek has launched a developer preview of Harness, an agent framework that operates on MIT license and has its source code available on GitHub. The framework's claim of "everything is a plugin" suggests a more tangible reality than what is typically presented in marketing. Harness is based on the Cordis kernel, a plugin kernel that manages tasks such as mounting, unmounting, and resolving dependencies.

All agent capabilities, including models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI, are plugins that can be composed through configuration files rather than forking the core code. This modular design allows for easy swapping of various components, such as model providers, sandbox implementations, or loop strategies, without affecting the main system.

Harness supports four runtime modes: standard, code, minimal, and creator. The standard mode includes a full suite of tools like file editing, shell execution, web search, planning, and workflows. The code mode enables the model to write TypeScript programs that can orchestrate multi-step operations in a single turn, following a similar pattern to Aider's Architect mode and Claude Code's thinking tool.

Minimal mode limits the framework to two tools, persistent bash and str_replace_editor, designed specifically for benchmarking purposes. Creator mode provides additional features for developing new agent configurations, including runtime inspection and preset authoring. Harness generates an append-only session log with system prompts, reasoning traces, tool calls, and results, as well as subagent scheduling and context injections.

This detailed logging enables empirical evaluation by allowing users to inspect, resume, fork, search, and replay events from the same event stream. The traceability of Harness could prove crucial in measuring the performance of different models on tasks like coding, as it provides a clear record of what each model observed and accomplished, rather than just the final output.

The plugin architecture holds promise for reproducible benchmarking, as it allows for the control of variables while swapping in different models. However, the exact influence of the Cordis kernel on behavior remains unverified, and the creator suggests running a minimal mode test to address this gap. The source code for Harness can be accessed at github.com/deepseek-ai/deepseek-harness, with the web interface available via the command npx @deepseek-ai/dsh web.