DeepSeek's Innovative Harness Treats Everything as a Plugin
The Chinese AI lab introduces a dynamic plugin system for machine learning models.
1Core Feature of DeepSeek Harness
DeepSeek has captured the developer community's attention by releasing an early version of its open-source AI harness. This harness is built on the Cordis meta-framework and treats everything as a plugin. Models, tools, skills, sessions, sandboxes, file systems, loops, orchestration, and user interfaces are all implemented as plugins. This allows developers to dynamically add, replace, and extend components.
2What is a Harness and Why Does It Matter?
A harness is a middleware layer that manages inputs passed to AI models and outputs returned from them. It handles prompt management, context management, tool orchestration, state management, and error handling. Harnesses like Claude Code and Codex perform similar functions for Anthropic and OpenAI models, respectively. Harnesses can significantly impact model performance and cost, making this a competitive field.
3Advantages of the Plugin System
DeepSeek Harness uses Cordis's plugin system to enable dynamic addition and removal of components without breaking the application. This provides developers with flexibility. Plugins cover everything from models to tools and user interfaces. Cordis's services and events ensure plugins work together seamlessly.
4Temporal and Spatial Composability
DeepSeek researchers highlight Cordis's temporal and spatial composability features. Temporal composability allows the effects of a component to be reverted upon removal. Spatial composability manages dependencies between components. These features prevent forced restarts and crashes in continuously evolving systems.
5Chain of Thought Traces
DeepSeek Harness records everything the model sees in a session log, including system prompts, reasoning, tool calls, subagent scheduling, and context injections. Developers can inspect these records in the Trajectory view and replay them. This is a useful tool for evaluating the model's reasoning process.
6DeepSeek R1 and Chain of Thought
DeepSeek R1 made waves last year with its chain of thought reasoning feature. This breaks down prompts into a series of 'thoughts' and reflects on these steps before producing a final answer. Access to intermediate reasoning is useful for assessing whether the model is reasoning correctly. Companies like Anthropic have started summarizing chain of thought traces to hide model reasoning.
7Harnesses as a Competitive Field
AI harnesses have become a competitive field, especially as models proliferate and become commoditized. Harnesses shape user habits by implementing the user interface. Once developers become accustomed to a harness, switching to a competing product becomes more difficult. Additionally, harness design can significantly impact model performance and cost.
8DeepSeek's Innovative Approach
DeepSeek Harness stands out with its innovative design, showcasing Chinese AI labs competing beyond model benchmarks and pricing. Treating everything as a plugin offers developers flexibility and customization. This could be a significant step forward for the future development of AI models.