Shortly after announcing API price adjustments, DeepSeek officially launched its first Agent product - DeepSeek Harness v0.1 Developer Preview on the evening of August 13th, with the project open-sourced on GitHub under the MIT license. This Agent product, which Liang Wenfeng has long valued, does not target the model itself, but rather the entire infrastructure that enables AI to actually perform tasks outside of the model. The official summarized its positioning in a concise formula: Model + Harness = Agent, meaning the model is responsible for thinking and reasoning, while Harness handles actual execution.
Harness includes all the necessary capabilities for a local agent workspace, including project management, long-term task collaboration, multi-agent orchestration, context management, online retrieval, and external skill calls. This means developers can obtain a base for running complex tasks directly without having to piece together tools from scratch. As an early preview version, its core design philosophy is summarized as "Everything is a plugin," aiming to redefine the way Agents are built through an open architecture.
Four Modes of Operation, Configuration as Assembly
The biggest highlight of Harness lies in its plugin-based open architecture: all capabilities such as models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and UI are composed of plugins, which can be freely replaced and flexibly reassembled. The entire framework is built on the Cordis plugin meta-framework, which only handles low-level tasks such as plugin loading and unloading and dependency management. All business components of Harness are independent Cordis plugins that collaborate through service and event mechanisms. Developers can freely assemble them at the configuration level without modifying the source code.
For different scenarios, Harness predefines four modes of operation and automatically loads corresponding plugin sets: the standard mode comes with a full set of tools to meet general development needs; the PTC programmatic tool call mode allows the model to generate code and orchestrate multi-turn tool chains; the minimal mode retains only Shell and file editing as two tools, used for baseline testing in minimal environments; and the creative mode allows viewing runtime status, memory debugging plugins, and customizing new modes of operation. After deploying a Node.js toolchain locally, developers can quickly launch a Web interface using the npx command.
As an open-source foundation for DeepSeek's layout in the Agent arena, v0.1 is still in an early stage, but its idea of being "disassemble-able and reconfigurable" addresses the current fragmentation in the intelligent agent ecosystem. At a time when OpenAI is promoting the Agent Plugins standard and various companies are competing for plugin influence, DeepSeek directly opens up the entire stack under the MIT license, lowering the barrier for developers to build their own Agents, and also making a significant contribution to the domestic open-source Agent ecosystem.
