Building a Custom AI Agent with DeepSeek Harness: A Hands‑On Validation

After reviewing DeepSeek Harness (DSH) documentation, the author runs dsh web locally and issues the instruction “Create a demand‑management menu that supports Markdown syntax for editing requirement documents.” The command triggers a long‑running DeepSeek v4 flash task, resulting in a new “Demand Management” entry in the left‑hand menu.

The menu opens to a full‑featured document management interface where users can create new documents and sort them by update time. The author demonstrates the Markdown editor with live preview and save functionality.

Resource usage for this task is displayed: 227 API calls, approximately 31.7 million tokens, costing ¥3.38.

According to official documentation, DSH serves both as a directly runnable coding agent (available in Web and headless modes, with extensibility via external profiles and plugins) and as an agent development framework. The demand‑management menu was created with a single sentence in the Web mode, illustrating the framework’s flexibility.

While DSH’s current coding‑agent experience lags behind mature solutions like Claude Code or Codex—due to its early stage, evolving APIs, and nascent plugin ecosystem—its true value lies in the framework component.

Core design: everything is a plugin . Models, tools, file systems, shells, sandboxes, session storage, sub‑agents, UI, and even the agent loop itself are implemented as plugins, akin to a LEGO car where the engine, wheels, and windshield can be swapped.

The author’s experiment confirms the simplicity of plugin implementation: the UI is a plugin, allowing the agent to modify its own interface.

Towards self‑evolving agents . DSH enables an agent to inspect its runtime, write a plugin on the fly, and mount it for subsequent tasks. This feature is experimental; dynamically generated plugins exist only in memory and are lost on restart, and they cannot yet be persisted as permanent plugins, though the language design makes this feasible.

Envisioning the future, the author suggests that once this capability matures, agents could instantly add missing functions through brief interactions, or even detect capability gaps during execution and develop, install, and invoke the required plugins autonomously.

As the ecosystem grows, a variety of high‑quality plugins are expected to emerge. Although DSH is still early, its “everything is a plugin” philosophy provides concrete space for community contributions.

Goal: Enable personalized AI agents that adapt to specific business scenarios.