DeepSeek's Super-Agent Doesn't Wait for Your Prompt, and That Changes What You Build

Over the weekend, DeepSeek quietly published something worth arguing about. Its Harness team open-sourced Super-Agent, an MIT-licensed prototype coding agent that, instead of waiting for you to type a prompt, monitors developer conversations in GitHub and Feishu, spots coding problems on its own, works on them in separate sessions, and drops the results back into the conversation, usually as a pull request. Bloomberg framed it as a direct shot at Claude Code, weeks before Anthropic's IPO, and that framing is accurate but incomplete.

Image: DeepSeek, via DeepSeek

The team's own words are the most interesting part. The goal, per the user manual, is to "take as little human attention as possible, and give as much value as possible." They argue that in traditional coding agents the bottleneck is not model capability, it is human attention, because one developer can only babysit one or two sessions at a time. Super-Agent is designed to break that limit by running work in the background and surfacing only finished results.

What the coverage missed

Almost every story I read about this falls into one of two buckets. Bucket one: geopolitics, a Chinese lab gunning for Anthropic right before its IPO, coming on the heels of US agencies being warned away from DeepSeek over data concerns. Bucket two: benchmarks and pricing, a free open-source alternative entering a market where Claude Code reportedly passes $2.5 billion in run-rate revenue.

Both buckets miss the part that changes your job. The interaction model just changed. Every agent you have used so far is a command box: you speak, it works, it stops. Super-Agent is ambient. It watches, it decides, it acts, and then it tells you what it did. That is a fundamentally different product to design, and almost nobody is ready for it.

My position: attention is the interface now

I think proactive agents are going to win, and the reason is simple. The people shipping with AI right now are bottlenecked not by typing speed but by attention. A solo designer running a studio can supervise three or four parallel builds if every step needs review. If the agent can own the long, boring stretches and show up with a finished diff, one person can run ten. DeepSeek is not even first with this idea, but open-sourcing it at zero cost is how an interaction model stops being an experiment and becomes the default.

Here is the part I want designers to internalize: once the agent acts without being asked, the product stops being a chat window and becomes a colleague. Colleagues are judged on behavior, not intelligence. When do you interrupt me, how do you prove what you did while I was away, how easily can I undo it, and how do you admit failure? Those are design questions, not model questions. The most important screens of the next five years are not prompts. They are the notification that says "I started something," the activity ledger that says "here is exactly what changed," and the undo button that says "we're fine, take it back."

There is a trust layer too, and I will not pretend it away. An agent from a lab that the US government told its own agencies to stop using will face questions about where your code context goes. That skepticism is healthy and it is one more reason the working prototype matters: it is open source, so teams can self-host, audit, and modify the behavior. Ironically, the transparency that makes it trustworthy has to be designed, communicated, and made visible in the UI. Even trust is a design deliverable.

What to do about it

If you build products with agents in them, and by now that includes almost every app worth building, treat this as your checklist:

1. Write presence rules before features. Decide explicitly when your agent is allowed to interrupt a user, and hold it to that. An agent that interrupts carelessly trains users to ignore it, which is worse than having no agent.
2. Design the receipt, not just the result. When work happens unsolicited, the user needs an artifact: what was the request, what changed, what files, what tests passed. Super-Agent returns pull requests, which is exactly the right instinct. The diff IS the interface.
3. Make undo a first-class action. Reversible agent actions get trusted faster. If everything your agent does can be taken back in one tap, users will let it do far more without watching.
4. Show failure modes honestly. DeepSeek's own docs admit the public prototype differs from their internal version and some described designs are not implemented yet. Respect to them for saying so. Products that hide what the agent cannot do get caught by users doing exactly the wrong task.
5. Budget attention like a resource. Track interruptions per session the way you track API spend. Your agent's quality will be judged in minutes of human time consumed, not tokens.

The takeaway

The command box was a phase. The next phase is agents that treat your attention as the scarcest resource in the building, and the winning products will be the ones that protect it deliberately. Designers and builders, that is not a threat, that is your lane. The person who decides how an ambient agent behaves, explains itself, and earns trust is doing product design at the highest level.

Want to build something with an agent in it this month? I still think Base44 is the fastest way for a designer to ship a real app around an idea like this, and its #1 slot in my AI builder ranking is exactly because it gets the human loop right.

The agent era is not about better answers anymore. It is about better questions, asked at the right time, by a teammate that knows when to stay quiet.