Can DeepSeek Harness Overtake VS Code? A Deep Dive into Its Agent‑First Architecture

VS Code’s extension ecosystem and Cursor fork

VS Code’s strength lies in its extension mechanism, which supports language servers, debuggers, Git integration, remote development, cloud services, and many efficiency tools. Cursor, a leading challenger, is a fork of VS Code’s open‑source code. It inherits the mature editing experience, language servers, and the large extension marketplace, then embeds AI deeply into code completion, editing, and agent workflows. Cursor’s approach assumes the editor remains the central hub for developer tools.

DeepSeek Harness architecture – “Everything is a Plugin”

DeepSeek defines itself as an “Agent Harness”, a layer that connects models to real work environments. Its architecture follows the slogan “Everything is a Plugin”. The official documentation lists the following plugin categories:

Model adapter

Tool registry

Session log

Filesystem and sandbox

Agent Loop

Sub‑Agent and workflow

Web UI

There is no privileged core component. At runtime the dsh process assembles a plugin tree, and different composition templates produce either a Web UI or a headless UI. Replacing any plugin—model, tool, permission policy, execution environment, or UI—yields a different product without changing the underlying runtime.

Why agents may shift control away from editors

Historically developers worked around an editor because code required manual inspection, modification, execution, and debugging. VS Code unified these actions in a single workspace and attracted an ecosystem via extensions, defining the previous developer‑tool era. Agents now can search files, modify code, run commands, handle errors, split tasks, and delegate work to sub‑agents. This shift raises new concerns that traditional editors do not address:

Accurate context acquisition for agents

Tool access control for models

Persistence and recovery of long‑running tasks

Collaboration among multiple agents

File‑ and command‑level permission management

Replaceability of models, tools, and execution environments

Integration of the same task across Web, desktop, or automation interfaces

DeepSeek Harness is built to solve these problems, moving the control layer from the editor to the agent runtime.

Cordis framework and dynamic plugin generation

DeepSeek Harness is built on the Cordis framework, which emphasizes composable and reversible plugins. When a plugin is unloaded, its services, events, and side effects are reclaimed. On top of Cordis, the platform provides dynamic tools that let an agent inspect the current runtime (available services, events, UI slots) and generate host‑side or browser‑side plugin code, mounting new capabilities on the fly.

Community members have generated plugins for session titles, work‑tree management, cost and balance displays, etc., without needing to study the full source code. This dynamic plugin ability is not enabled by default; developers must activate it explicitly, and the generated plugins do not survive a dsh restart, indicating an experimental status.

Microsoft’s independent Agent Host

Microsoft is developing a separate Agent Host that detaches long‑running agent sessions from the VS Code extension process. The Agent Host can operate without an editor client, connect to remote workspaces, and host multiple agent runtimes, addressing the limitation that traditional extensions are tied to the editor window’s lifecycle.

Running DeepSeek Harness and the DSH Desktop client

The official way to start DeepSeek Harness is a single Node.js command: npx @deepseek-ai/dsh web To lower the entry barrier for users without Node.js, the author open‑sourced DSH Desktop, a macOS/Windows client that bundles Node.js 24 LTS, allows installing or switching official @deepseek-ai/dsh versions, and runs the official Web UI in an independent window. The desktop client does not modify DeepSeek Harness’s core code; it only manages the Node environment, version selection, and local process, while the original DSH continues to handle models, sessions, plugins, skills, and user data.

Repository:

https://github.com/qufei1993/dsh-desktop

Current limitations (Developer Preview)

DeepSeek Harness is in Developer Preview and warns of future incompatible changes. It lacks a full code‑editing workspace, debugging ecosystem, source‑control experience, and a large extension marketplace. The Web UI focuses on agent task interfaces; code reading, diff, and command output are still maturing.

Community feedback is split. Some users praise the openness and ability to generate custom plugins. Others report slower performance, high token consumption, and a steep learning curve for plugins, profiles, and Cordis configuration. Sub‑agents, documentation, and certain interactions also exhibit early‑stage issues.

Outlook

If the agent‑centric platform matures, the control layer may move from the editor to the agent runtime, making editors a possible UI surface rather than the central hub of the developer ecosystem.