Here is a new article crafted from the provided post content, now including an FAQ section and a conclusion section.
—
# DeepSeek Harness v0.1: A Developer Preview of a Plugin-Based Agent Runtime
DeepSeek has released **DeepSeek Harness v0.1** as a developer preview, distributed under the MIT license and available as `dsh` at [deepseek-ai/deepseek-harness](https://github.com/deepseek-ai/deepseek-harness). At its core, Harness is a layer—the “harness”—sitting between an AI model and the environment it interacts with. This layer provides the tools, files, sandboxes, and control loop that enable an agent to operate continuously.
DeepSeek frames this architecture simply as **Agent = Model + Harness**. Unlike many existing agents where the loop, tool registry, and session store are hard-coded, Harness takes a “plug everything” approach. As stated in its README, *everything is a plugin*. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and even the UI all sit behind Cordis plugin boundaries. This makes Harness less a fixed coding assistant and more a kit for assembling agent runtimes.
## The Cordis Kernel
Harness runs on **Cordis**, a meta-framework designed around spatiotemporal composability. The kernel handles plugin mounting, unmounting, and dependency resolution. Capabilities live in the plugins, not in a privileged core. Developers can select, swap, or extend any capability—such as models, tools, or scheduling—via configuration without ever modifying the Harness source code.
## Four Runtime Modes
Harness ships with four distinct runtime modes, each loading a different default set of plugins:
– **Standard**: A full coding agent featuring file editing, shell access, file and web search, skills, planning, goals, subagents, and workflows.
– **Code Mode**: Exposes all Standard tools through a Code Mode SDK, allowing models to combine multi-step operations into a single TypeScript program.
– **Minimal**: A stripped-down environment with only two persistent tools—`bash` and `str_replace_editor`—designed primarily for benchmarking model performance.
– **Creator**: Enables runtime inspection, in-memory plugin experimentation, and preset-authoring guidance for building custom agent configurations.
## Traceable Execution
Every run in Harness is fully traceable. All inputs and outputs—including system prompts, reasoning, tool calls and results, subagent scheduling, and context injections—are written to an append-only session log. This log supports operations such as resume, fork, search, and replay, all operating on the same event stream. While many agent frameworks log tool calls, Harness emphasizes recording every context injection.
## Model Routing as a Plugin
Model routing is treated as a plugin as well. Adding providers such as Anthropic or OpenAI via API key takes effect immediately without restarting the server. Native-auth providers like Bedrock, Vertex, Azure, and Codex require respective credentials. Custom providers can connect to any OpenAI-compatible endpoint. API keys are stored write-only in `$DSH_HOME/.credentials.yaml`, with settings retaining only credential references.
## Running DeepSeek Harness
To start the Web UI, developers can run:
“`bash
npx @deepseek-ai/dsh web
“`
Alternatively, from a cloned repository:
“`bash
git clone https://github.com/deepseek-ai/deepseek-harness
pnpm install
pnpm run build
pnpm dsh web
“`
A Python SDK is also available as `deepseek-harness-sdk`, requiring Python 3.10+ and running on Linux x64, Linux arm64, or macOS 14+ on arm64.
—
## FAQ
**Q: Who is DeepSeek Harness v0.1 intended for?**
A: It is primarily intended for AI-native startups, platform and developer-experience teams within mid-to-large enterprises, regulated enterprises running local pilots, and specific industries such as software development, financial services, healthcare R&D, cloud and semiconductor vendors, and academic research labs.
**Q: What are the main applications of Harness?**
A: Use cases include internal coding agents over private repositories, model evaluation in controlled two-tool environments, agent observability and run replay, custom sandbox and approval policies, and packaging house tooling as reusable plugins.
**Q: Is Harness deployable in production?**
A: Yes, but it is positioned as developer infrastructure rather than a finished end-user product. It is self-hosted and MIT-licensed, making it suitable for piloting at the company level.
**Q: How are models managed in Harness?**
A: Model routing is managed through a plugin system accessible via a settings UI. API keys are stored securely and applied dynamically. The system supports multiple providers and allows custom OpenAI-compatible endpoints.
**Q: What makes Harness different from other agent frameworks?**
A: Harness treats almost every component as a swappable plugin, emphasizes traceability through an append-only session log, and offers multiple runtime modes for different use cases. It is designed as a runtime kit rather than a fixed assistant.
—
## Conclusion
DeepSeek Harness v0.1 represents a shift in how agent runtimes can be composed and extended. By treating everything as a plugin and running on the Cordis kernel, it offers exceptional flexibility for developers building AI agent workflows. With four distinct runtime modes, full traceability, and provider-agnostic design, Harness is positioned as powerful developer infrastructure for those serious about building and experimenting with agentic systems. As a developer preview, it invites exploration and collaboration from startups, enterprises, and research communities alike.
