Back to list
Product LaunchDeepSeekAI AgentsOpen Source

DeepSeek Launches Harness Developer Preview: A Fully Modular Plugin-Based Framework for AI Agent Development

DeepSeek has officially released the developer preview of DeepSeek Harness, an open-source framework designed to bridge the gap between AI models and real-world environments. Built on the Cordis kernel, the platform adopts an "everything is a plugin" philosophy, allowing developers to swap or recompose every capability—from models and tools to UI and scheduling—via configuration. With a focus on traceability, DeepSeek Harness records every interaction in an append-only session log, providing a transparent "Trajectory view" of agent reasoning and actions. This release includes source code and developer documentation, aiming to empower the global community to build highly customizable and reliable AI agents by providing a robust infrastructure that manages dependencies and environment interactions.

Hacker News

Key Takeaways

  • Modular Architecture: DeepSeek Harness is built on the "everything is a plugin" principle, allowing for the swapping and recomposition of models, tools, skills, and UI components.
  • Cordis Kernel Integration: The framework utilizes the Cordis kernel to manage plugin mounting, unmounting, and complex dependencies between agent capabilities.
  • Full Traceability: Every action, including reasoning, tool calls, and context injections, is recorded in an append-only session log for complete transparency.
  • Developer-Centric Design: The framework is open-source and supports configuration-based customization, enabling developers to extend capabilities without modifying the core source code.

In-Depth Analysis

The "Everything is a Plugin" Philosophy

At the core of DeepSeek Harness lies a radical commitment to modularity. Unlike monolithic agent frameworks, DeepSeek Harness treats every functional component as a swappable plugin. This includes the fundamental building blocks of an AI agent: the models themselves, the tools they use, their specific skills, and even the underlying infrastructure like storage, sandboxes, and scheduling loops.

By building on the Cordis kernel, DeepSeek Harness provides a sophisticated management system for these plugins. The kernel handles the mounting and unmounting of capabilities, ensuring that dependencies are resolved correctly. This architecture allows developers to select, swap, or extend any capability through simple configuration changes. This means that a developer can update a model or change a UI component without ever touching the primary DeepSeek Harness source code, significantly lowering the barrier to iterative development and experimentation.

Bridging the Gap: Agent = Model + Harness

DeepSeek defines the modern AI agent through a simple yet profound formula: Agent = Model + Harness. In this paradigm, the model serves as the "soul" of the agent—the central intelligence capable of reasoning and processing information. However, a model in isolation cannot effectively function in the real world.

The "Harness" acts as the essential interface that allows the model to understand its environment, interact with external tools, and maintain persistence across real-world settings. By providing the necessary sandboxes, sessions, and scheduling loops, the Harness ensures that the agent's intelligence is grounded in a functional context. This distinction emphasizes that while the model provides the reasoning, the Harness provides the utility and reliability required for production-grade AI applications.

Traceability and the Trajectory View

One of the most critical features of the DeepSeek Harness developer preview is its focus on traceability. In the development of complex AI agents, understanding why a model made a specific decision or how a tool call failed is paramount. DeepSeek Harness addresses this by implementing an append-only session log.

This log records every single element the model encounters or produces, including:

  • System prompts and context injections
  • Internal reasoning processes
  • Tool calls and their subsequent results
  • Subagent scheduling and task management

This data is visualized through the "Trajectory view," providing developers with a clear audit trail of the agent's behavior. This level of transparency is vital for debugging, auditing, and refining agent performance, ensuring that every run is fully traceable and reproducible.

Industry Impact

The release of DeepSeek Harness as an open-source developer preview marks a significant step toward the standardization of AI agent infrastructure. By providing a modular framework that separates the "soul" (model) from the "body" (harness), DeepSeek is offering a blueprint for how complex AI systems can be built with greater flexibility and reliability.

The use of a plugin-based system built on the Cordis kernel suggests a move toward a more collaborative ecosystem where developers can contribute and share specific agent capabilities. Furthermore, the emphasis on traceability addresses one of the primary concerns in AI deployment: the "black box" nature of agentic workflows. By making every step of an agent's trajectory visible and loggable, DeepSeek is setting a high bar for transparency in the industry, which is essential for building trust in autonomous AI systems.

Frequently Asked Questions

Question: How can I start using DeepSeek Harness?

Developers can begin by installing the framework via npx using the command $ npx @deepseek-ai/dsh web or by cloning the repository directly from GitHub at https://github.com/deepseek-ai/deepseek-harness. Detailed developer documentation and quick-start guides are available to assist with the initial setup.

Question: What is the role of the Cordis kernel in this framework?

The Cordis kernel serves as the management layer for the entire plugin system. It is responsible for mounting and unmounting plugins, managing their dependencies, and facilitating communication between different services and events. This allows the various capabilities of the agent to work together seamlessly.

Question: What components can be customized in DeepSeek Harness?

Virtually every aspect of the harness is customizable. This includes the AI models, tools, skills, session management, sandboxes, storage solutions, execution loops, scheduling logic, and the user interface. All of these can be swapped or recomposed through configuration files.

Related News

Gemini API Tool Retrieval Preview Gives Agents Access to Hundreds of Tools With Fewer Tokens
Product Launch

Gemini API Tool Retrieval Preview Gives Agents Access to Hundreds of Tools With Fewer Tokens

Google AI Blog announced that Tool Retrieval is now in preview in the Gemini API. The feature is presented as a way to give agents access to hundreds of tools while using fewer tokens. The announcement pairs those two points: the breadth of tool access and lower token use. It does not explain how Tool Retrieval works, specify how token use was measured, or quantify the reduction. It also does not provide an exact tool count beyond describing access to hundreds of tools. The stated release status is preview, and the announcement does not give a timeline for general availability.

LangSmith Tool Gateway Enters Public Beta for Secure, Governed AI Agent Tool Use
Product Launch

LangSmith Tool Gateway Enters Public Beta for Secure, Governed AI Agent Tool Use

LangChain announced that LangSmith Tool Gateway is in Public Beta. The gateway handles credentials, identity, and access for every agent tool call, with policies, approvals, and traces. The announcement frames secure, governed tool use around the individual call: its focus is on how agents access tools, rather than on what those tools or agents can do. The update identifies what the gateway covers but does not explain how credentials are managed, how access decisions are made, when approvals apply, or what information traces contain. The Public Beta is the only release status stated in the announcement.

Google’s Gemini Decision 1 Picks Answers Instead of Writing Them Across Five Input Types
Product Launch

Google’s Gemini Decision 1 Picks Answers Instead of Writing Them Across Five Input Types

Google’s Gemini Decision 1 is a model designed to pick answers rather than write them, according to Shittu Olumide’s report for KDnuggets. It reads text, images, video, audio, and PDFs in a single forward pass and supports up to 262,144 input tokens per request. The central distinction is between what the model reads and what it does with that input: it accepts several kinds of material, but its stated task is answer selection, not answer writing. The token figure describes the maximum input per request, not how much the model outputs. The report does not explain how answers are presented for selection or provide results showing how well the model performs.