{
  "id": 926150,
  "title": "DeepSeek's innovative harness treats everything as a plug-in",
  "url": "https://urgent.news/2026/08/14/deepseeks-innovative-harness-treats-everything-as-a-plug-in-926150",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-14T20:25:56.000Z",
  "source": {
    "name": "The Register Science",
    "slug": "the-register-science",
    "url": "https://www.theregister.com/ai-and-ml/2026/08/14/deepseeks-innovative-harness-treats-everything-as-a-plug-in/5288095"
  },
  "original_language": "en",
  "account": "DeepSeek has captivated the developer community with the early release of its open source agent harness. This innovation follows the growing importance of harnesses in the realm of machine learning models. Engineered on the Cordis meta-framework, the DeepSeek Harness operates on the fundamental principle that all components are plugins. Models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, and UI are all implemented as plugins, allowing for seamless mixing, matching, replacement, and extension.\n\nThe term \"harness\" has gained prominence this year, referring to middleware or a mediation layer that manages input to an AI model and output. Harnesses regulate prompts, context management, tool orchestration, the agent loop, state management, error handling, safety, permissions, and related concerns. Claude Code handles the harness for Anthropic's Claude model family, while Codex serves a similar function for OpenAI's GPT model family, among others like Aider, Cline, Goose, OpenCode, OpenHands, and Pi.\n\nThe term's vagueness means it can refer to the agent loop and tools or extend to broader concerns such as sandboxing, subagents, and more. AI model harnesses are now key battlegrounds in the AI competition, particularly as models become abundant and standardized. Harnesses often manage user interfaces, creating user inertia and making it challenging to switch products once developers adapt to a particular interface, even if it's primarily a command line. Moreover, studies indicate that model performance and cost can vary significantly depending on the harness used, due to diverse design choices.\n\nFor instance, DeepSeek Harness relies on a minimal system prompt of around 200 tokens, while Claude Code utilizes a system prompt of approximately 10,000 tokens. The same model yields different results with different harnesses. DeepSeek Harness stands out due to its innovative design and as a demonstration of Chinese AI labs aiming to compete beyond model benchmarks and pricing. The harness treats everything as a plugin, utilizing the plugin system from its underlying Cordis framework. This framework enables dynamic addition and removal of components without disrupting the application, a concept known as temporal and spatial composability. DeepSeek Harness also records every interaction, allowing users to inspect, resume, fork, search, and replay events. This feature sets it apart from competitors, offering a more comprehensive and flexible approach to AI model management.",
  "summary": "Chinese AI labs keep moving forward while US labs play defense",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "The Register",
        "title": "DeepSeek's innovative harness treats everything as a plug-in",
        "url": "https://urgent.news/2026/08/14/deepseeks-innovative-harness-treats-everything-as-a-plug-in",
        "published": "2026-08-14T20:25:56.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}