Playbooks · Coding tasks
DeepSeek's Super-Agent is a prototype — treat it that way
Written November 12, 2026
Most agent demos ask you to trust the system a little more. This one's own README starts by asking you to trust it less: research code, not a product, may contain bugs, review and adapt it before it goes anywhere near your work. What the prototype does is still worth understanding — a coding agent that lives in one GitHub repository and answers requests with pull requests.
The warning is the headline
The repository, published by DeepSeek's Harness team on October 31, 2026 under the MIT licence, opens with its status in bold: a research prototype for reference only, not an official DeepSeek product. The code grew out of the team's day-to-day development work, much of the engineering was done by coding agents, and the style and practices may be inconsistent. The public version is not identical to DeepSeek's internal version or to the system described in its design article, Exploring Super-agent Design, and some designs in that article are future directions that are not implemented in the prototype. There is no model announcement, price or product date in the repository, because there is no product.
How it works
The system, called dsh-superagent, is a proactive coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness. It listens to GitHub and, optionally, to Feishu, the workplace chat app. When someone allowed asks for work, the agent opens a work item, does the job in its own agent session and its own directory, and answers in the original conversation — for example with a pull request. A web panel shows what needs a person. Setup is itself agent-led: you ask your coding agent to read the repository's SETUP.md and work through it with you, one short question at a time, with the full plan shown once before anything is installed.
The design rule: your attention is the budget
The repository states its working principle directly: take as little human attention as possible, and give as much value as possible. An agent that can act for a whole team can also waste a whole team's attention, so this one acts only when asked, or when a rule the team wrote allows it, and speaks only when someone must act or can use a result. The numbered documents in the repo walk through the philosophy, the domain model, authorization, execution and deployment in that order — authorization gets its own document, which is the right emphasis for a system designed to act without being watched minute by minute.
Limits
- One repository per deployment, GitHub plus optional Feishu. Anything broader is outside the published design.
- The prototype status is not modesty. DeepSeek says the code may still contain bugs and is not production-ready; running it against a repository that matters, without reading it first, ignores the only instruction the authors put in bold.
- Parts of the design article describe where the team wants to go, not what the code does today. Anyone evaluating the idea should read the repository's documents, not the coverage of the article.
Related playbooks
- DeepSeek, Exa, Brave, Tavily: which web-search API should your agent actually pay for now that search is metered
- Audit apps that installed agent plugins for you
- BotPsy 0.8B: an agent model that fits on-device
- Budget DeepSeek Search and Fetch calls
Sources
Back to: Coding tasks playbooks hub · Also: Kimi Code 2.5.3: steer a running session · Kimi Code 2.6.0 canary: read the label