SAN FRANCISCO—DeepSeek released its agent harness framework on Aug. 13 in developer preview, publishing the full source code under the MIT license. The release arrived alongside the company's V4 Pro model and is available to agent framework developers worldwide.
The architecture's central design decision: every agent capability is a plugin. Models, tools, skills, sessions, sandboxes, storage, agent loops, scheduling and the user interface are all discrete, swappable units. A developer can replace any one of them in configuration without touching the underlying DeepSeek Harness source code.
The plugin system runs on a kernel called Cordis, which handles mounting, unmounting and dependency resolution between plugins. Cordis services and events serve as the communication layer that lets plugins operate together. The Cordis design is documented in a paper titled "A Programming Paradigm for Spatiotemporal Composability."
DeepSeek Harness ships with four distinct operating modes. Standard mode provides the full tool set. Code mode uses model-generated code to coordinate multiple rounds of tool calls rather than routing through a static tool registry. Minimal mode strips the environment down to a shell and a file editor, making it usable for clean benchmarking runs where researchers want to isolate model performance. Creator mode lets developers inspect the live runtime, test Cordis plugins in memory and combine them into new custom modes on the fly.
Every interaction the model has is written to an append-only session log. The log captures system prompts, the model's reasoning steps, tool calls and their results, subagent scheduling records and every context injection. DeepSeek calls this the Trajectory view. Developers can inspect those records filtered by source, then resume a session, fork it into a parallel branch, search across the event stream or replay any prior state. All four operations draw from the same underlying log.
Installation requires Node.js. Developers can launch the web interface with a single npx command, or clone the full repository and follow the setup instructions for a local build. The MIT license imposes no commercial restrictions, meaning enterprises can use, modify and redistribute the harness without licensing fees.
The MIT license choice is the business story here. DeepSeek is a Chinese AI research lab that first drew wide attention in January 2025 when it published weights for its R1 reasoning model, also under MIT. That release forced a rapid recalibration across the AI industry because R1's benchmark performance approached frontier U.S. models at a fraction of reported training cost. The Harness release extends the same open-source strategy into the infrastructure layer one level above the model itself.
The competitive pressure that creates for U.S. agent framework companies is direct. LangChain, the dominant Python-based agent orchestration library, and LlamaIndex both charge for hosted versions and enterprise support. Microsoft's Semantic Kernel and AutoGen are open-source but backed by Azure's commercial cloud. DeepSeek Harness enters that field with no licensing cost, an architecture where the model adapter is itself a plugin—meaning it is not locked to DeepSeek's own models—and a session logging design that is more auditable than most comparable frameworks.
The plugin-first approach also matters at the infrastructure economics level. Enterprise AI teams currently face a fragmentation problem: they may run different models for different tasks—a small model for tool-call routing, a larger one for reasoning, a fine-tuned model for domain-specific retrieval—across providers that include Anthropic, OpenAI, Google and open-weight alternatives. A harness where the model adapter is a swappable plugin and the agent loop is itself a plugin means a team can reassign workloads across providers in configuration alone, without rewriting orchestration code. That lowers switching costs, which is precisely the kind of architectural decision that large enterprise buyers evaluate when assessing vendor lock-in.
DeepSeek Harness remains in developer preview. The project is hosted at the deepseek-ai organization on GitHub under the repository name deepseek-harness. The API surface and plugin contracts are subject to change before a stable release—a practical caution for teams considering production deployments now.
DeepSeek has now open-sourced at the model layer, the reasoning layer and the agent infrastructure layer within roughly 19 months. Each release has arrived with a permissive license and documentation that lowers the barrier for enterprise adoption. The cumulative effect is a vertically integrated open-source AI stack that a company can run entirely on-premises with no dependency on a U.S. cloud vendor—a selling point that carries weight in markets where data sovereignty requirements restrict cloud data transfers.
