DeepSeek's Free AI Super Agent: We Gave It a Real GitHub Project and 5 Issues

You've got five AI chats open. One wants approval. One finished ten minutes ago and you missed it. DeepSeek's new free "super agent" is built for exactly that problem: it isn't built to write better code than Claude, it's built to bother you less.

Everyone's calling it a Claude Code killer. So we gave it a real project and five GitHub issues. Two of them were traps. It got three clean passes and two with a catch, for 11 cents in total. And we found a bug nobody had mentioned.

Why this matters even if you never use GitHub

DeepSeek measured how many AI sessions one person can really follow at once. The answer is three to five. The AI isn't the slow part anymore. You are.

So the interesting question isn't "can it code?" It's "does it know when to leave you alone?" That's what our test was built to check.

What it actually is

In plain words, it's an AI team member that lives on one project. Someone on an approved list writes "@agent, fix this." It opens a work ticket, does the work in its own private session, and replies with a pull request. A pull request is a proposed change that a person reviews before it goes in.

It's free and open source under the MIT license. It's a plugin for DeepSeek Harness, and DeepSeek calls it a research prototype, not an official product. The AI model behind it costs a few cents.

Every request goes through a check before any work starts. The check asks two things: am I allowed to ask, and is this really a request? If both are yes, it opens a ticket with a written permission that says exactly how far the work can go.

Test 1: a real bug

Our demo project is a bill-splitting app with a planted bug. A $120 dinner, four people, 20% tip. The app says $31.50 each. It should be $36. And every test in the project still passes.

The check said yes and yes. Twenty-nine seconds later it replied "on it." Fifty-two seconds in, there was a pull request. It found the real mistake: the tip was being divided twice. It fixed it in two lines and added three tests so it can't come back. That cost 1.6 cents.

Test 2: vague feedback

Next, we tagged it and wrote, "the split page looks kind of off." No target. No real request.

Most AI tools would jump in and redesign your page. This one read it and did nothing. No reply. No code. Here's the note it left for itself: "vague visual feedback, no target, direction is unsettled."

That's its first rule: it acts only when it's asked, or when a rule your team wrote allows it. The second rule: it speaks only when you have to act, or when there's a result you can use.

Test 3: a decision it should have asked about

This is where it got interesting. We asked it to add currency support. Sounds simple, except someone has to decide which currencies, and whether a group can mix them. That's a product decision, not a code fix.

By its own rules, it should stop and ask. It didn't. About two minutes later we had a currency picker with twelve currencies. Euros written the European way. Yen with no decimals. Before-and-after screenshots it took of the page by itself.

It's good work. But it made the decision for us: labels only, no conversion. To be fair, it told us that in plain words.

The catch is in its own design. When you tag it directly, it skips the check that's supposed to catch design decisions. So here's the rule of thumb: if you tag it, you're the one who already decided.

Test 4: a follow-up with no tag

We went back to its pull request from test 1 and left a normal review comment: please show it as dollars and cents. We didn't tag the agent at all.

Twelve seconds later: "on it." Two and a half minutes later: $38.33. Correct.

But look at the files. It left seven leftovers in the pull request: screenshots, and copies of code it used to check its own work. A real reviewer would send that back.

Still, this is the idea that matters. Once a ticket exists, every later reply on that thread goes straight to that ticket. Nobody has to tag the AI again.

DeepSeek's own numbers (and the fine print)

DeepSeek tested the super agent against one normal AI session handling about three hundred messages. The super agent finished even its slowest jobs (the 95th percentile) in 15.7 minutes. The single session took 29.5. And it never handed work back to a person: zero times out of five.

But here's the fine print from DeepSeek's own write-up: the coding itself was simulated by a timer. So their test proves it routes work faster. It says nothing about whether the code is any good. That's why you need a hands-on test like ours.

Test 5: a stranger asks, and safety

For the last test, we took ourselves off its approved list and asked for a new button like a stranger would. Twenty-nine seconds later, it ignored us. Silently. The reason it logged: "not a member."

Good, because this thing runs with full access to your computer. DeepSeek says so in its own docs, and tells you to give it a separate user account on your computer with only the agent's passwords. Anything nobody asked for starts in observe mode: it writes down what it would have done, and does nothing.

Scoreboard

Three clean passes, two with a catch. The whole test cost 11 cents on DeepSeek V4.1 Flash through OpenRouter.

Versus Claude Code

We gave the same bug to Claude Code running Opus 5.5. It fixed it in 19 seconds, on our machine, while we watched. No pull request, and we had to be there. Claude Code is the faster worker. DeepSeek built the coworker that picks up work from a comment while you're somewhere else.

The catch: the lost messages

There's one more thing we didn't plan to find. Twice, we tagged it, and it simply never saw the message. No error. Nothing. That's 2 of our 6 first mentions.

The reason is in the timing. GitHub told the agent about a mention about 23 seconds late, and the agent only looks back 5 seconds. So a first mention can just vanish. It's the kind of thing you only find by actually running it.

Our honest read: the judgment is impressive. The plumbing is still a prototype, exactly like DeepSeek said.

Your questions answered

• Do you need DeepSeek Harness? Yes, it's a plugin for it.
• Do you need DeepSeek's own servers? No. We ran it on DeepSeek's Flash model through OpenRouter.
• Do you need Feishu, the chat app? No. GitHub alone works.
• Is setup hard? You tell your coding agent to read SETUP.md and set it up with you, and it asks one question at a time. The one step it can't do for you is make a second GitHub account for the agent.
• Should a beginner run it on real work? Not yet. It's research code. But you can borrow its rules today.

Sources: the dsh-superagent-prototype repo on GitHub (MIT, research prototype) and DeepSeek's Harness article on X from November 1, 2026.

Get the Quiet Agent Kit

We turned the super agent's rules into a short instruction block you can paste into Claude Code or ChatGPT. Your AI checks in once, stays quiet, and only comes back when you need to act. The kit also has the setup map and our five test issues word for word, so you can run the same test yourself. Comment QUIET on the video, or grab it below.

How we tested: a private demo project where we wrote every message, five GitHub issues, run November 3, 2026, one run each, real billed cost on DeepSeek V4.1 Flash via OpenRouter. It ran under our normal computer account, not a separate one. One small test, not a study. Independent, not affiliated with DeepSeek or Anthropic.