Overview
- DeepSeek published the Super-agent prototype as MIT‑licensed research code on Oct. 31 and said the public repository is a reference implementation that differs from its internal system.
- The prototype monitors a single GitHub repository and optional Feishu chats, turns approved requests into work items, routes them through a persistent Router session into separate Worker sessions, and returns results as pull requests or items for a local web panel.
- Project documentation warns that Router and Worker sessions run with the full access of the operator's OS account and recommends using a dedicated OS account and starting in observation mode to limit credential exposure.
- DeepSeek reported an internal experiment where parallelized agents completed a task in under 16 minutes versus more than 29 minutes for a conventional approach, but the company’s benchmark and the public code’s performance are not independently verified.
- The release drew rapid attention on GitHub and forms part of DeepSeek’s wider strategy, supported by roughly 100 billion yuan in fundraising ahead of a planned IPO, to challenge Anthropic and OpenAI in coding agent tooling.