DeepSeek built an AI boss that hands out work to other AIs

DeepSeek has published a long write-up on what the next generation of AI agents should look like and, at the same time, open-sourced a working prototype on GitHub under the MIT license. The idea is called Super-agent, and it goes straight at the biggest pain of anyone living in Claude Code, Codex and similar tools.

Today an agent lives inside a single session. Want to run five tasks, keep five windows open and remember what is where. DeepSeek's team figures a person can really track three, maybe five active sessions before it turns into a mess. And an agent on your laptop goes to sleep when the laptop does.

The fix: a Router sits on top and acts as the boss. It takes in every message, keeps the shared context and hands out work. Below it, Workers run in parallel, each in its own isolated session. Only the boss talks to the user, so it feels like one assistant even when a whole crew is grinding behind it.

The agent listens to more than a chat box. Team chat messages, emails, calendar reminders, GitHub code comments and CI results all become events. Plain rules filter them first, then the model decides whether to act at all. Once a task belongs to a specific Worker, new comments on it go straight to that Worker, skipping the boss.

The fun part is speculative execution. A meeting invite arrives, and the agent checks your calendar, gathers background and drafts a reply in advance. Colleagues argue about a bug in chat, and the agent quietly tries to reproduce it, only showing up once it has found a simple fix.

Compute and state can live in the cloud while commands run on your own machine through a lightweight daemon. A separate Watchdog session watches for breakage and for Workers drifting off task.

There are numbers, but they come from the lab. On a synthetic load of about 300 messages, roughly 20 of them real work requests, Super-agent finished tasks in 12.4 minutes at the median and 15.7 minutes at the 95th percentile. A standard single session setup with a neutral prompt hit 29.5 minutes at the 95th percentile. The cost is more cached tokens read, and DeepSeek itself notes the test did not measure actual code quality.

Inside the company this agent already works: it reproduces and fixes bugs, responds to code reviews and keeps improving DeepSeek Harness itself. The public dsh-superagent prototype watches one GitHub repository and, optionally, Feishu, picks up a task when asked and answers in the same thread, for example with a ready pull request. Setup is unusual too: you ask your own coding agent to read SETUP.md and configure everything with you.

DeepSeek's own caveat: this is research code, not a product, some ideas from the write-up are not in the prototype yet, and the security of a cloud agent reaching into a local machine is still an open problem.

Verdict: China just showed how to turn a chatbot into a coworker that finds its own work, and gave away the blueprints for free.