01What Is DeepSeek Super Agent on GitHub?
DeepSeek Super Agent GitHub refers to DeepSeek's open source AI agent team that lives on GitHub. One lead agent takes your request, hands the work to worker agents, and they build it and send the finished result back.
Here is how it works in plain terms:
- Super Agent lives on GitHub.
- It runs inside DeepSeek Harness.
- You ask it for something on a GitHub issue.
- The lead agent replies and starts workers.
- The workers send back finished pull requests.
Julian Goldie set it up on a demo business, a dental practice, and built six things with it. He ran the whole team on DeepSeek V4.1 Flash, and he shows each build running.
DeepSeek describes Super Agent as a research prototype. It is open source under the MIT license, so you can read the code and run it yourself.
The idea behind the tool is to use as little of your attention as possible, and only speak up when someone needs to act.
02How to Set Up DeepSeek Super Agent
Super Agent runs as a bundle inside DeepSeek Harness. Before you start, you need these tools on your machine:
- Node.
- pnpm.
- Git.
- tmux.
- Python.
- The GitHub command line tool.
Then follow these steps:
- Give the agent its own GitHub account.
- Add that account to your repository.
- Put your own name on the allowed list.
- Pick your model. Julian used DeepSeek V4.1 Flash.
- Choose a mode: observe mode or auto mode.
- Open the panel on your own machine.
The separate account matters. The agent ignores its own comments, so the person asking and the agent must be different accounts.
The two modes work like this. Observe mode only records what the agent would do on its own. Auto mode lets it act on the rules you have written. Starting in observe mode is a safe way to see what it would do before you let it act.
The panel shows every ticket, every worker and anything that needs you, all in one place.
03Build 1: A Jarvis That Talks to the Agent Team
The first build connects a voice assistant to the agent team. Julian uses his own Jarvis dashboard, running locally on his Mac.
Here is the flow:
- He tells Jarvis to ask the agent team to add a map and directions section to the homepage.
- Jarvis opens an issue on GitHub and tags the agent.
- Jarvis tells him it has handed the job over.
- About a minute and 40 seconds later, the lead agent replies that it is on it.
- Jarvis reads the reply out loud.
Just before that, Julian had asked Jarvis for a sticky call now button for mobile visitors. That came back as a finished pull request in two minutes and 42 seconds. The map section he asked for on camera landed as a pull request 10 minutes after he asked.
So the whole loop runs by voice. You talk to Jarvis, Jarvis files the issue, the agent team builds, and you get a pull request to review. You never open the code yourself.
04Build 2: A Premium Animated Website
Next, Julian asked the agent team to rebuild the dental site as a premium animated site. Here is what came back:
- A slowly rotating golden pearl gem in 3D that follows your mouse.
- Big, elegant type.
- Sections that glide in as you scroll.
- Glass service cards that glow when you hover.
- A booking form that checks what you type.
It is still one single file with no build step, so it loads fast. It even has a calm version for people who switch off motion on their phone.
The best part is how changes work. When Julian wanted something different, he did not touch the code. He left a comment on the pull request, and the worker came back and made the change.
That is a simple way to work for anyone who is not a developer:
- Ask for the site on an issue.
- Review the pull request.
- Leave a comment with what you want changed.
- Let the worker update it.
05Build 3: A Website Audit Tool for Finding Leads
The third build is a website audit tool for finding leads. You give it a list of local business websites, and it checks each one and scores it out of 100.
It checks the things that lose a business enquiries:
- Is the site secure?
- Does it work on a phone?
- Is there a proper page title and description for Google?
- How fast does it load?
- How heavy is the page?
The results go into a dark animated dashboard, ranked. Then the tool writes a short email for every site, naming its top fixes.
Julian ran it on seven pages: versions of his demo site plus one public test page. Seven sites were scored and seven emails were drafted in under a second.
If you sell websites or SEO, this gives you a list of prospects with the reasons to contact them already written. You still need to read the emails before you send them, but the research is done for you.
06Build 4: An SEO Agent Swarm for Landing Pages
The fourth build is an SEO agent swarm. Julian filed three requests, one page each: Invisalign, teeth whitening and emergency dentist, all for the local area.
Here is what happened:
- The lead agent answered all three requests within about a minute.
- Three workers started at the same time, each in its own copy of the code.
- The first page came back in under four minutes.
- All three were done in under seven minutes.
Each one is a full landing page with:
- A clear headline.
- What the treatment is.
- Who it is for.
- What happens at the appointment.
- A frequently asked questions section.
- Structured data, so Google understands it is a dental practice.
The pages also link back to the homepage and to each other. Julian's tip is that you could queue a whole list of these before bed and wake up to a stack of pull requests to review.
07Build 5 and 6: A Front Desk Bot and Self-Fixing Reports
Build 5 is a front desk that never closes. Julian asked for a chat widget that answers common questions and takes booking requests. Six minutes later, the worker sent back 22 files, including the answers, the booking steps, tests and nine screenshots proving it works.
In the demo, he asks what time they are open on Saturday, and it answers. He asks about parking, and it answers. He says he would like to book, and it walks him through name, phone number, day and time.
That request lands in a front desk inbox, so the team can call the patient back.
Build 6 fixes things nobody asked it to. Julian filed a normal problem report saying the phone number on the homepage was wrong. He did not tag the agent or ask it to do anything.
- Super Agent watches the repository and saw the report.
- It matched the report to one of its rules about copy fixes.
- It decided this was its job and replied that it was on it.
- Two minutes after the report, there was a pull request.
One line changed: the old number out and the new number in, with before and after screenshots. It also said nothing else changed.
08Who DeepSeek Super Agent Is For
DeepSeek Super Agent suits business owners and builders who are happy to work through GitHub issues and pull requests, and want a team of agents doing the building.
It fits best if you:
- Run a website and want changes made from a simple request.
- Sell websites or SEO and want audits and outreach emails drafted.
- Need several landing pages built at once.
- Want a chat widget that takes booking requests.
- Like the idea of small fixes happening without being asked.
Keep these limits in mind:
- It is a research prototype, not a finished product.
- It needs several developer tools installed first.
- It needs its own GitHub account, separate from yours.
- Start in observe mode before you switch to auto mode.
Julian mentions that every request from this episode, the Jarvis agent team handoff, the step-by-step Super Agent setup and the faceless video skill are inside the AI Profit Boardroom, along with a course on agent teams from zero.