DeepSeek publishes the design for an agent that works without being asked

What DeepSeek published

As qbitai.com reported, DeepSeek released an engineering write-up on a "Super-agent" and a prototype repository on GitHub. The article lays out how a personal AI agent could keep working across several tools without being told what to do each time.

The headline claim is that, inside DeepSeek's own team, the agent noticed a bug, tracked it down and proposed a fix (a pull request, meaning a suggested code change) without anyone prompting it.

The problem it targets

Developers jump between a team chat, where a plan is discussed, and GitHub, where the code is reviewed. A traditional coding agent treats each as a separate session, so the person has to carry the context between them. DeepSeek says a person can usually follow only 3 to 5 active sessions at a time.

In the internal example, a team chose approach A in chat, then asked for approach B in a GitHub comment. The agent recognized both messages as one job and changed course.

A manager, workers and a watchdog

The top layer, the Router, is a dispatcher that takes in messages and assigns work. Below it, Workers each handle one task, in parallel. A Worker that picks up a pull request can subscribe to its comments, so later feedback goes straight to it.

For guessing ahead, the agent can use speculative execution: trying something in the background before anyone asks. If a bug comes up in chat, it tries to reproduce it and only speaks up if it finds a workable fix. A separate watchdog AI looks for crashes and for tasks stuck in loops or drifting off target.

What the test showed, and what it doesn't

DeepSeek ran a simulated workplace with a stream of chat messages and a batch of work requests, repeating each setup several times. Its slow-end completion times were shorter than those of a traditional single-session setup. Removing the direct routing of messages made the Router do more work and read more cached data, which raises cost.

These are DeepSeek's own preliminary results on a simulation. We have not seen evidence that real maintainers accepted the agent's pull requests, so treat the bug-fix story as a demonstration, not a track record.