Research prototype open-sourced by the DeepSeek Harness team 10-31 under deepseek-ai: an autonomous coding agent dedicated to a single GitHub repository, packaged as a DSH plugin. It monitors GitHub (optionally Feishu/Lark); when an authorized user @-mentions the agent account with a request, it creates a ticket, executes the task in its own session and directory, and returns the result (e.g., a pull request) to the original discussion thread. The local web dashboard lists only items requiring human intervention.
Star count 2490
Confidence: High (100) · Event time 2026-10-31
Rationale
- New official developments within the past 24 hours: Repository 10-31 created (independent media reports cite 09:32 UTC), MIT license; on the same day, the Harness team published an article on their official WeChat account titled "Super-agent Design Exploration: Towards More Autonomous Agent Systems." According to Yicai Global, internal instances are already involved in daily issue reproduction and fixing, code review responses, and the iteration of DeepSeek Harness itself (the original WeChat article could not be accessed for this column).
- The architecture follows a "one main session for dispatching + one execution session per ticket" model: The main session uses the dsh-superagent-router preset to determine whether to accept a task. Each ticket spawns a worker-<ticket-number> session and an independent directory. Multiple tickets can run in parallel with a configurable limit. Foreground commands switch to the background after 5 seconds, and background tasks can run for up to 30 minutes after the turn ends.
- Design principles are documented in docs/000: Minimize human attention consumption. Only direct @-mentions from authorized users count as requests. Proactive work not explicitly requested requires team-defined rules, and defaults to "observe" mode, which only logs "what would have been done" without actually executing it. The agent remains silent during execution, replying only when results are ready, decisions are needed, or tasks cannot be completed.
Usage
- Prerequisites: Node.js 22.19+ or 24+, pnpm, git 2.32+, tmux, Python 3.8+, GitHub CLI 2.48+. DSH must be within the engines range specified in package.json (deployment documentation pins npm install -g @deepseek-ai/[email protected]), and DSH must be configured with a model service.
- Official recommendation: Let the coding agent guide you through installation. After cloning the repository, tell the agent, "Read SETUP.md in this repository and set up dsh-superagent with me." It will ask one question at a time (repository, agent account, who can assign tasks, whether to auto-merge, etc.), present a complete plan, and then execute it. Manual steps are documented in docs/006-deployment.md (uses a separate superagent profile; dashboard defaults to 127.0.0.1:3082).
- After installation, use an authorized account to comment @<agent-account> <request> in a repository issue. A ticket should appear on the dashboard within approximately one minute.
Risks
- The README explicitly states this is a research prototype, not an official DeepSeek product: The code is largely generated by coding agents and may contain defects. It differs from internal versions and the implementation described in the article; some designs mentioned in the article are future directions and not implemented in the prototype.
- Execution sessions run commands with danger-full-access permissions using the runner's own system account without requiring approval. It can execute programs by path and read your home directory. Documentation recommends isolating it using a separate system account containing only agent credentials. The agent must be a dedicated GitHub user account (GitHub Apps are not supported). Only github.com is supported; GitHub Enterprise Server is not. Feishu/Lark requires creating a separate bot application and lark-cli.
- Star count (2490 at time of collection) only reflects attention level. There are currently no public benchmarks or usage data. Internal usage details are based on the WeChat article and media reports.