DeepSeek's Super-Agent, Explained: A Router Dispatches, Workers Execute, and It Speaks Only When Allowed

What happened

The prototype splits a repository's day-to-day collaboration into two layers: a resident router takes conversations, evaluates requests and dispatches work, while isolated workers execute in their own sessions and directories and send results back to the original conversation, for example as a pull request. Each approved request becomes a ticket carrying an immutable authorization statement, and later comments on a pull request return to the same worker without a fresh mention. A companion web panel lists only the items that need a person.

The companion article, Exploring Super-agent Design: Towards a More Autonomous Agent System, describes the internal use: according to MadRobot's account of the article, DeepSeek's internal super-agent already reproduces and fixes bugs, responds to code reviews and helps build DeepSeek Harness, reading public group-chat feedback, locating root causes and submitting fix pull requests without being mentioned. The article's argument is an attention bottleneck: according to MarkTechPost, a person can actively follow only 3 to 5 sessions at once, so attention, not model capability, constrains parallel work.

Authorization rules are the shared gate for the prototype and the internal version. The documentation defines three accept modes, off, observe and auto; in the default observe mode the agent only records requests it would accept and does not execute them, and team-written rules plus requests from approved people are the two channels that open a ticket. Behaviour during a run is tightened the same way: confirm once, stay quiet while working, and speak again only with a result, a question or a failure.

For background, Harness is the open-source agent framework DeepSeek released in August 2026, also under the MIT license; the super-agent prototype ships as a bundle for Harness, runs on the operator's own machine, and needs a dedicated Harness profile plus Node.js 22.19 or later, Git, Python, tmux and the GitHub CLI. Several reports say the repository collected thousands of stars within a day of going up, and several reports relay Bloomberg's description of the release as a challenge to Anthropic's Claude Code.

What they said

DeepSeek draws the line up front in the repository statement: the prototype is for reference only, not an official DeepSeek product and not production-ready code, it may still contain bugs and its engineering style may be inconsistent, so users should review and adapt it before use. The statement also says the published code was adapted for release and is not identical to the team's internal version or to the implementation in the design article, and some designs in the article are future directions not implemented in the prototype. The README states the working principle verbatim: take as little human attention as possible, and give as much value as possible.

MarkTechPost relays the design article's criticism of the single-session model: mainstream agents treat one session as the basic unit, people can follow only a limited number of sessions in parallel, and attention becomes the bottleneck, so the super-agent's router-worker split exists to spend attention on items that must be handled by a person. The same report notes that the simulation figures are self-reported by the team, not an independent evaluation.

Third-party reports converge on the limits. MarkTechPost notes that the prototype runs on the operator's own machine under a dedicated Harness profile, with sessions holding full access as the operating-system user, so a separate operating-system account is needed for isolation; CompleteAITraining adds that the current version supports only github.com, not GitHub Enterprise Server. Neither report describes the prototype as a ready-to-use product.

Key facts

1. Publisher and date: DeepSeek's Harness team, published on GitHub on Oct 31, 2026; repository deepseek-ai/dsh-superagent-prototype, created Oct 31, 2026 at 09:32 UTC, MIT license, TypeScript bundle, version v0.1.0, a single commit at release
2. Positioning: a proactive coding agent for a single GitHub repository, packaged as a bundle for DeepSeek Harness; listens to GitHub and optionally Feishu
3. Architecture: a resident router dispatches work, isolated workers execute; each approved request becomes a ticket with an immutable authorization statement; the web panel lists only items needing a person
4. Authorization modes: off, observe and auto, with observe the default, recording but not executing; team rules and approved requests are the channels that open a ticket
5. Simulation: about 300 messages and about 20 work requests, each scheme run 5 times; median completion 12.4 minutes, 95th percentile 15.7 minutes against 29.5 minutes for the single-session baseline; about 23.2 million cached tokens per run against 6.1 to 10.8 million for the baseline; figures come from the team's own simulation, not an independent evaluation
6. Requirements: the operator's own machine with a dedicated Harness profile; Node.js 22.19 or later, Git 2.32 or later, Python 3.8 or later, tmux and GitHub CLI 2.48 or later; github.com only
7. Limits: a research prototype for reference only, not an official product and not production-ready; the published code differs from the internal version and the design article's implementation, and some designs are future directions