DeepSeek wants its AI to stop waiting for instructions. The Hangzhou lab's Harness team announced on its official WeChat account Saturday evening that it has open-sourced Super-agent, a prototype coding agent that sits in a team's GitHub repository and work chat and can pick up a problem — and propose a fix — before anyone asks it to, according to Elets CIO.
The contrast with today's tools is deliberate. Claude Code's GitHub integration acts when someone tags @claude; OpenAI's Codex works a session you opened. DeepSeek's internal instance of Super-agent, per the team's design essay, reads feedback in public group chats, works out the root cause of a reported problem, and submits a pull request with a fix, unprompted.
The shipped prototype is narrower: a "proactive coding agent for one GitHub repository" bundled for DeepSeek Harness, the lab's open-source rival to Claude Code. A permanent Router handles every conversation and dispatches jobs; isolated Worker sessions carry them out in parallel, each in its own directory. In DeepSeek's own simulation the multi-session design cut 95th-percentile completion time from 29.5 to 15.7 minutes under noisy message traffic — though the test used synthetic messages, simulated tools, and didn't measure code quality.
What makes this release interesting isn't the autonomy, it's the leash. Unprompted work is off by default: the agent ships in observe mode, recording what it would have done, and acts on its own only when a team's written rules allow it. The guiding principle is to "take as little human attention as possible, and give as much value as possible" — and DeepSeek adds the blunt advice teams probably need: "Do not use it to produce volume."
That caution reads as the real thesis. DeepSeek argues agents are now held back less by models than by people — a person can only track three or four active agent sessions at once, so parallelism is capped by human attention, not model capability. The industry has been racing to make agents smarter; this is a race to make them quieter. It's the same idea OpenAI was circling with the always-on "dots" agents it showed at DevDay in September, now released as inspectable MIT-licensed code.
DeepSeek is careful not to oversell it: the README calls it "research code, not production-ready software," says much of it was written by coding agents, and notes some essay ideas — including an external watchdog — aren't implemented yet. Operators are told to run it under a separate OS account. That warning lands with extra weight one day after we covered a tampered Claude Code extension that planted an SSH backdoor on developers' machines.
The context around the release is loud too. The Harness team was formed this year specifically to build agents that compete with American coding products; its Harness repo has passed 188,000 GitHub stars; and Bloomberg reports DeepSeek has raised about $15 billion ahead of a planned IPO. The company isn't just competing on model weights anymore. It's trying to own the workflow the models run inside.
— via Elets CIO