DeepSeek released a developer preview of Harness, an open source framework for building AI agents in which every capability is a plugin that can be swapped or recomposed, according to DeepSeek's documentation. The swappable pieces include models, tools, skills, sessions, sandboxes, storage, loops, scheduling and the user interface.
The framework runs on what DeepSeek calls a Cordis kernel, which manages plugin dependencies so developers can "select, swap, or extend any capability in configuration without changing the DeepSeek Harness source code," the documentation says.
Every action a model takes, system prompts, reasoning, tool calls and results, subagent scheduling and context injections, is recorded to an append-only session log, according to DeepSeek. The framework ships with four runtime modes: a standard mode with the full toolset, a code mode for model-generated orchestration, a minimal mode for benchmarking, and a creator mode for building custom presets.
Harness is released under the MIT license and available now through the command `npx @deepseek-ai/dsh web` or by installing from GitHub, where the companion deepseek-ai/deepseek-harness repository passed 85,000 stars within a day of release, according to GitHub. DeepSeek said core plugins and APIs will keep evolving during the preview.
For teams building their own agent harness rather than adopting a vendor's, an append-only log of every model decision is the auditability piece most homegrown frameworks skip until something breaks in production.