What is DeepSeek Super Agent?
DS just dropped a Super Agent prototype on github:
It is a prototype coding agent under their harness bundle:
Most headlines call it a Claude Code rival. That misses the point. The interesting part is not what it can do. It is what it tries not to do: waste your team's attention.
What it is
The project is called dsh-superagent. It is a coding agent that works on one GitHub repository. It listens to GitHub and, if you want, Feishu chat. When someone on the allowed list asks it for work, it opens a ticket, does the work in its own session and folder, and answers in the same conversation, usually with a pull request. A small web panel shows only the things that need a person.
Be clear on what it is not. DeepSeek calls it a research prototype, not an official product. They say much of the code was written by coding agents. It is an experiment made public. Treat it that way.
The philosophy
The whole project rests on one rule: take as little human attention as possible, and give as much value as possible. Every message, notification, question, setting and screen element costs someone's attention. If it does not give back more than it costs, it goes.
That rule shapes how the agent behaves. It acts only when asked, or when a rule the team wrote allows it. It speaks only when a person needs to act or can use the result. It asks one clear question at a time. It stops based on impact, not effort. A small fix goes all the way to a pull request. A change that needs a design call stops and asks.
It also puts the burden on whoever runs it. Start narrow. Tune for less noise. Don't use it to pump out volume. Own what it says.
We like this. Most agent tools compete on how much they can do. This one competes on how little it bothers you.
How it differs from Claude Code and Codex
Claude Code and Codex are tools you sit with. You open a session, give it a task, watch it, steer it, close it. The developer is in the loop the whole time.
Super Agent flips that. It is a service, not a session. It runs in the background and takes requests from your team on GitHub. The unit of work is a ticket, not a session. A ticket is created, tracked and closed on its own, whether anyone is watching or not.
Under the hood there are two roles. A long running Router handles incoming messages and hands out tasks. Isolated Workers do the actual coding, each in its own space. Think receptionist and specialists. When someone comments on a pull request later, the comment goes straight to the Worker that owns it. No need to tag the agent again.
The other big difference is trust. Every ticket carries a record of who asked and why the work is allowed. And proactive work, where the agent spots something nobody asked for, starts safe. There are three modes: off, observe and auto. Observe is the default. The agent only writes down what it would have done. You watch that list before you let it act on its own.
To be fair, Claude Code and Codex can also be tagged on GitHub. What is new here is the always on service that sorts through a stream of chat and CI messages, the observe first approach to proactive work, and the attention budget built into the design.
The benefits
More work in parallel. Each request gets its own worker, so tasks don't wait in line. In DeepSeek's tests, the median task took 12.4 minutes and the slowest 5% took 15.7 minutes. A single session agent took 29.5 minutes for that same slowest 5%.
Less babysitting. Nobody keeps a terminal open. You ask in the thread where the work already lives, and the answer comes back there.
Safer autonomy. Allowed lists, written rules and observe mode let you widen what the agent does step by step, not all at once.
It's open. MIT license. You can read it, change it and run it on your own machine. DeepSeek Harness also works with other model providers, not just DeepSeek. For teams in regulated industries, that matters a lot.
One caveat on the numbers. DeepSeek tested with fake messages and simulated tools. They did not measure the quality of the code. Speed is shown. Quality is not, yet.
Who benefits most
Small engineering teams with one main repo and a steady stream of small, clear tasks. Bug fixes, missing tests, dependency bumps. The kind of work that piles up because nobody wants to stop what they're doing.
You need decent CI, so the agent's pull requests get checked before a human looks. And you need one person willing to own it: set the rules, read what it posts, fix the prompts when it gets things wrong.
Who should wait? Solo developers, who will get more out of Claude Code or Codex in the terminal. Teams spread across many repos, since it works on one. And anyone who can't isolate it. It runs with full access to the user account on the machine. Give it its own machine or VM and its own GitHub account.
What it costs to run
The software is free. You pay for three things.
Model tokens. This is the main bill. On DeepSeek's own API, the Flash model costs $0.15 per million input tokens and $0.60 per million output tokens off peak. Peak prices are double. The bigger V4 Pro model costs $0.66 in and $1.98 out off peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Everything else is off peak, including weekends. For a team in North America, almost the whole workday is off peak.
Caching matters even more. Cached input on Flash costs $0.003 per million tokens off peak. An agent that keeps rereading the same repo and the same instructions benefits a lot from that. If you use another model provider instead, your costs follow their prices.
A machine. It runs as a service, so it needs somewhere to live. A small VM is enough to start.
Attention. This is the real cost, and the project says so itself. Someone sets it up, writes the rules and reviews every pull request it opens. If that person isn't there, the cheap tokens don't matter.
Our take
Even if you never install it, steal the idea. In our client work, agents rarely fail because the model isn't smart enough. They fail because they are noisy, they ask too much, and people stop trusting them. An agent judged by how little attention it costs is an agent people will keep using.
Super Agent is a prototype. But it points at where coding agents are going: less chatbot, more teammate with a narrow job.
If you're thinking about putting agents into your team's workflow, especially in a regulated industry, talk to us. That's the work we do every day at Softmax.