Defining the 'Agent-Centric Android Era'? DeepSeek Launches Harness Framework

On the 13th, DeepSeek officially launched the developer preview of its first AI agent (Agent) development framework, 'DeepSeek Harness' (hereinafter DSH). Simultaneously, it introduced the DeepSeek-V4-Pro model, optimized specifically for agent performance. The project is open-sourced under the MIT license and gained over 50,000 GitHub stars within just 12 hours of release, demonstrating significant interest from the developer community.

The 'Everything-as-Plugin' Modular Design

The core design philosophy of DSH is 'everything is a plugin.' Built on the Cordis plugin system, the framework modularizes every component—models, tools, sandboxes, user interfaces (UI), and even the agent's main execution loop (Agent Loop)—as plugins.

DeepSeek officially likens the model to the 'brain' and Harness to the 'hands and feet,' emphasizing that the value of an agent lies in task completion, not just conversational output.

Project lead Cui Tianyi noted that the current 0.1.0 version is an early-stage developer preview, with interfaces and the plugin ecosystem rapidly evolving. He compared DSH to a 'super LEGO car,' explaining that the official coding agent is merely a default configuration. Developers can freely swap engines, tires, or add modules, even assembling entirely new forms.

The Potential and Debate Around 'Self-Evolving Software'

Alongside DSH, an 88-page technical paper was published, introducing the 'spatiotemporal composability' paradigm, aiming to enable agents to continuously generate and replace components without human intervention. Early internal tester Jiayuan (JY) Zhang believes DSH already shows early signs of 'self-evolving software,' where agents can write and mount plugins on the fly to gain new capabilities.

However, this feature has sparked industry debate. Developer Baoyu (@dotey) expressed reservations, suggesting 'software self-evolution' might be a false problem, as software requires rigorous design and validation—excessive automation could lead to chaos. He also pointed out that DSH currently has a high entry barrier, requiring Node.js toolchain knowledge, making it unfriendly to non-engineers.

Transparent Logging and a Shift in Business Strategy

In practical testing, DSH's 'Trajectory' feature received high praise. This functionality records the entire process—including system prompts, chain-of-thought reasoning, and tool calls—achieving full data traceability. In terms of cost, real-world tests showed token consumption under 3 RMB for tasks like webpage reconstruction and API calls, highlighting its economic efficiency.

Analysis suggests DeepSeek aims to emulate Android's open approach, shifting its competitive focus from merely selling tokens to building an agent ecosystem. With the launch of the V4 series models, DeepSeek also announced a peak/off-peak pricing strategy starting August 17, using off-peak discounts to guide resource allocation.

This framework, seen as a prototype of an 'Agent OS,' seeks to attract global developers to co-define the next generation of AI productivity standards by opening up 'shell-level' permissions.