Separate capabilities clearly
Use the plugin-first architecture to reason about model access, tools, and workflow responsibilities as separate parts.
DeepSeek Harness
Use DeepSeek Harness as a reference for composing your own agent workflows. Study its plugin architecture to separate model access, tools, and task capabilities, then evaluate focused extensions for work organized in Floatboat.
About DeepSeek Harness
DeepSeek Harness, also called dsh, is an open-source agent harness developed by DeepSeek. It provides a runtime around model interaction rather than a new model in itself. The project’s central design is to express the system through composable plugins.
Its value for workflow design is the separation of capabilities into composable parts. A Floatboat user can use that idea to identify the tools, context, and responsibilities a task requires before considering a specific plugin or runtime integration. The reference architecture does not establish a ready-made connection to Floatboat.
DeepSeek describes the project as a developer preview with compatibility-breaking changes expected. Its MIT source license is separate from model-provider access and third-party dependencies. Evaluate it as an evolving platform, with version-specific documentation and deliberate workspace boundaries.
Workflow Mechanisms to Learn From
Understand what the project demonstrates, then decide which mechanism addresses a real gap in your own workflow.
Compose a focused workflow
Evaluate only the capabilities needed by the task, rather than treating a large plugin catalog as a complete working solution.
Plan for extension changes
Treat version compatibility as part of workflow design so a runtime update does not silently change the expected behavior.
Apply the Ideas to Your Workflow
Use these scenarios as design references for work organized in Floatboat. Any software integration needs its own compatibility and permission checks.
A modular information workflow
Break collection, transformation, and review into explicit capabilities with clear inputs and outputs.
A custom tool connection
Study plugin contracts when an existing workflow needs a specific external tool or a task capability unavailable in the current setup.
A controlled workflow experiment
Evaluate a small composition and record its behavior before relying on a preview-stage runtime for recurring work.
From Reference to a Workflow Experiment
Start with a task and a small testable change, whether that means better instructions, a new tool connection, or a runtime component.
- 01
List the task capabilities
Define which information, tools, and intermediate results each stage needs, independently of a particular plugin name.
- 02
Check a candidate composition
Review capability boundaries, credentials, and version compatibility for the extensions you actually need.
- 03
Test the handoffs and failure path
Verify that a failed tool or changed plugin produces a visible result and a clear decision about the next step.
Related Agent Projects
Compare independent agents through their interfaces, extension models, and approach to carrying work forward.
DeepSeek Harness FAQ
Is DeepSeek Harness an official DeepSeek project?
Yes. It is published in the deepseek-ai organization, and its README explicitly identifies DeepSeek AI as the developer. Use that repository rather than assuming a similarly named community fork is official.
What does everything-is-a-plugin mean here?
It describes the project’s Cordis-based composition model. The runtime’s capabilities are assembled through plugins; the actual tools and behavior still depend on the plugins and configuration selected for a session.
Is dsh a separate AI model?
No. dsh is the runtime and application entry point. Model inference comes from the provider configured for the session, with that provider’s own credentials, availability, and commercial terms.
How stable is the project interface?
The official README marks it as a developer preview and explicitly expects compatibility-breaking changes. Pin and evaluate the version you use instead of assuming an extension will work unchanged across releases.
Does MIT cover connected models and services?
MIT applies to the project source. Model APIs, external services, and third-party components retain their own terms, costs, and operational requirements.
Must I switch products to use these workflow ideas?
No. You can first use DeepSeek Harness as a reference for task structure, tool boundaries, and result review in your existing workflow. Reusing software is a separate choice that requires compatibility, permissions, and license checks; applying the design ideas does not automatically connect the software to Floatboat.