DeepSeek dsh-superagent-prototype: a router-and-worker agent system that keeps working without being asked

DeepSeek's Harness team open-sourced a 'super-agent' prototype on Oct 31: a two-layer Router/Worker architecture with event-driven, asynchronous multi-channel interaction, compute-state separation, and an external watchdog, packaged as a bundle for DeepSeek Harness. It listens to GitHub and Feishu, manages background coding sessions, and answers with pull requests — an internal instance already reproduces bugs, answers code reviews, and iterates on Harness itself.

Why it matters

What you could build with it

Does it hold up?

A working research prototype, not a product: DeepSeek reports an internal instance running continuously on real bug reproduction and code review, but public use requires setting up DeepSeek Harness and the bundle, so production readiness is unproven outside DeepSeek.

Built with DeepSeek dsh-superagent-prototype

• DeepSeek builds 'super-agent' for autonomous codingtechinasia · Reports the Oct 31 WeChat technical post: a system that spots defects reported in group chats and tries to reproduce and fix them in the background without being tagged.
• DeepSeek's New Architecture Lets AI Start Work Without Being Askedtrendingonweibo · Reads the design as DeepSeek competing on agent scaffolding rather than model benchmarks, and flags the open reliability question: an agent that acts unprompted can also be wrong unprompted.
• DeepSeek Proposes Super-agent Architecture Design, Exploring More Autonomous Agent Systemspanews · Summarizes the five-part architecture — Router/Worker layers, event-driven design, human-oriented interaction, compute-state separation, and Watchdog — and the single-session limits it targets.
• deepseek-ai/deepseek-harnessgithub · 267,500 ★ · The plugin-oriented harness the prototype plugs into; the internal super-agent instance runs on it.

Learn more

technical deep dive →

First spotted on github: source.

More AI agents