DeepSeek Harness

Overview

DeepSeek Harness is an open-source agent development framework from DeepSeek. Built on the "everything is a plugin" philosophy, it breaks down every agent capability — models, tools, skills, sessions, sandboxes, storage, loops, scheduling, UI, and more — into independently replaceable plugins. Developers can freely select, swap, or extend any capability at the configuration layer without touching source code, effortlessly assembling agent applications tailored to real business scenarios.

The framework builds on the Cordis core, which only handles plugin loading, unloading, and dependency management, while services and events enable plugins to collaborate. Every run is recorded in an append-only session log, including system prompts, chain-of-thought, tool calls and results, sub-agent scheduling, and more. Developers can trace, resume, fork, and replay from any source in the Trajectory view, ensuring every run leaves an auditable trail.

Harness offers four presets: Standard mode for full-featured tasks like coding and web retrieval, PTC mode where models compose multi-step tool calls via a single code snippet, Minimal mode with only shell and file-editing tools for benchmark testing in constrained environments, and Creation mode for experimenting with plugins at runtime and crafting custom presets. The project is now open source under the MIT license and can be launched with a single npx command, making it easy for developers to get started and deeply customize.

Core capabilities

Plugin-based architecture

All agent capabilities are provided by independent plugins that can be freely swapped and recombined.

Cordis core

A lightweight core that only manages plugin loading and dependency resolution, not specific capabilities.

Trajectory replay

Append-only session logs capture the full run, supporting resume, fork, and retrieval.

Multi-mode support

Standard, Minimal, PTC, and Creation modes adapt to different scenarios.

Configuration-layer composition

Extend or replace any capability in config files without modifying source code.

Open-source MIT

Fully open code, free for commercial use and secondary development.

One-command launch

Run the Web UI quickly via an npx command with a low learning curve.

Use cases

Build your own coding agent: use Standard mode to let the agent autonomously write, test, and fix code

Multi-tool collaboration automation: combine search, file editing, and Shell at the configuration layer to automate complex tasks

Model benchmarking: use Minimal mode to evaluate a model's coding and tool-use abilities in a controlled environment

Custom agent presets: experiment with plugins in memory via Creation mode to craft new run modes

Auditable agent operations: trace every AI decision through session logs and the Trajectory view

Plugin ecosystem development: build and publish custom plugins to push the boundaries of Harness

Enterprise workflow orchestration: use scheduling and loop plugins to keep agents working in production