DeepSeek has rolled out version 0.2 of DeepSeek Harness (DSH), the company's open-source agent harness, in public preview. Announced on September 29, the update is the first to offer the framework as ready-to-run desktop installers, with builds for macOS on Apple silicon and 64-bit Windows available through the official Harness website.

The release arrives roughly seven weeks after the initial v0.1 developer preview, launched on August 13, and follows preview builds of an Electron-based app that were spotted circulating from DeepSeek's update server earlier in the week, ahead of any formal announcement.

Built on Cordis's "everything is a plugin" architecture, DeepSeek Harness can run either as a desktop application or as a web interface launched from code, with the latter requiring only Node.js — a setup that v0.2's installers make optional for the first time. The framework is positioned as a ready-to-use environment for running AI agents across several domains: everyday office work such as organizing files, analyzing data, drafting documents and creating presentations; software development tasks including repository exploration, bug fixing and test execution; research work with source verification; and background jobs such as batch processing.

DeepSeek describes Harness as the most widely used coding agent among users of its official model API, measured by daily active users and sessions, while acknowledging that the product remains at an early stage.

What's New: Plugin Management, Automation and Reworked File Handling

The centerpiece of v0.2 is a dedicated plugin installation and management page. Extending Harness previously required command-line work; users can now install plugins simply by entering an npm package name — a change that lowers the technical barrier to extending the agent — and can disable, uninstall and inspect each plugin's description and source from a single interface.
The update ships alongside six official plugins — Agent Teams, voice input, Shell, Agent Loop, Subagent and Web Search — and introduces an experimental Creation mode in which users describe a desired plugin in conversation and Harness writes or modifies it automatically. Newly created plugins take effect immediately and appear in the manager. DeepSeek reports that roughly 60% of Harness users run third-party plugins, a level of reliance on community-built extensions that contextualizes the company's plan to introduce an official plugin marketplace, together with a more stable plugin API designed to produce fewer breaking changes.

Version 0.2 also adds a built-in automation tasks plugin, enabled from the plugin page, that converts repetitive work into scheduled tasks complete with run history, adjustable frequency and editable instructions. According to a hands-on review by tech outlet ifanr, scheduled tasks continue running even after the application is closed. The plugin page lists four experimental features in total: agent teams, automatic authorization review, automation tasks and voice input — the last of which downloads a roughly 1 GB local speech-recognition package based on SenseVoice, keeping speech processing on the user's machine.

Beyond packaging, the release substantially reworks how files, content previews and code changes are presented. Users can hand Harness documents, spreadsheets or PDFs for analysis and chart generation, or request project modifications with code execution. Generated files and each round of code changes are collected within the conversation, a side panel previews files and diffs, and results can be opened in local applications. Users can also control how much of the agent's working process is visible.

Looking ahead, DeepSeek's roadmap includes a stronger sandbox with improved security, enhanced agent teams, personalized long-term memory, browser and GUI automation, remote and mobile access, and session sharing with multi-user collaboration. The company also indicated it will tune future models specifically for performance within the Harness environment.

Source: Metaverse Post