DeepSeek open-sources its team agent, and the design brief is about saving attention
DeepSeek just showed how it runs a coding agent inside its own team. On November 1 the DeepSeek Harness account published a long article, "Exploring Super-agent Design: Towards a More Autonomous Agent System", which DeepSeek's main account reposted and which passed 840,000 views within hours. Alongside it sits a repo, dsh-superagent-prototype, MIT licensed, already at about 7,000 stars and 700 forks.
The prototype is a coding agent that lives in one GitHub repository and, optionally, in Feishu group chats. Someone on the allowed list mentions it, it opens a work item, runs it in its own session and directory, and answers in the same thread, usually with a pull request. A small web panel lists only the things that need a person. Healthy states take no space.
The interesting part is the philosophy file. The whole system is judged by one rule: take as little human attention as possible, give as much value as possible. Every message, notification, setting and paragraph of docs has to pay for the attention it costs or it gets cut. The agent acts only when asked or when a rule the team wrote allows it, and proactive rules start in a record-only mode. It asks one concrete question at a time. The authorization doc is even stricter: messages, comments and quoted documents are inputs, never permission changes. Prompt injection gets handled as an architecture question, not a filter.
DeepSeek is careful to call it a research prototype, not a product, and says much of the engineering was done by coding agents. Setup itself is delegated: you tell your own coding agent to read SETUP.md and it walks you through. Bloomberg framed it as DeepSeek going after Anthropic in coding agents. That is true, but the bigger signal is the shape. The fight is moving from the agent in your IDE to the agent sitting in your team chat, and the scarce resource it spends is not tokens but your colleagues' attention.
Repo: https://github.com/deepseek-ai/dsh-superagent-prototype Article: https://x.com/DeepSeekHarness/status/2116857257245987236
← Back to all articles
The prototype is a coding agent that lives in one GitHub repository and, optionally, in Feishu group chats. Someone on the allowed list mentions it, it opens a work item, runs it in its own session and directory, and answers in the same thread, usually with a pull request. A small web panel lists only the things that need a person. Healthy states take no space.
The interesting part is the philosophy file. The whole system is judged by one rule: take as little human attention as possible, give as much value as possible. Every message, notification, setting and paragraph of docs has to pay for the attention it costs or it gets cut. The agent acts only when asked or when a rule the team wrote allows it, and proactive rules start in a record-only mode. It asks one concrete question at a time. The authorization doc is even stricter: messages, comments and quoted documents are inputs, never permission changes. Prompt injection gets handled as an architecture question, not a filter.
DeepSeek is careful to call it a research prototype, not a product, and says much of the engineering was done by coding agents. Setup itself is delegated: you tell your own coding agent to read SETUP.md and it walks you through. Bloomberg framed it as DeepSeek going after Anthropic in coding agents. That is true, but the bigger signal is the shape. The fight is moving from the agent in your IDE to the agent sitting in your team chat, and the scarce resource it spends is not tokens but your colleagues' attention.
Repo: https://github.com/deepseek-ai/dsh-superagent-prototype Article: https://x.com/DeepSeekHarness/status/2116857257245987236
Comments