JoeBot demonstrates Manus, an AI agent that autonomously tests a web app by spidering the DOM and running exploratory scenarios without page objects or specs. It finds a critical 404 error on a call-to-action button, generates a bug report, feeds it to Cursor to auto-fix the code, then re-runs tests to confirm the fix works. The demo warns against letting the same AI both write fixes and verify them.
Can AI autonomously find bugs on a website, fix the code, and then verify the patch, all without human intervention? Well, in this video, I put Manas and Cursor to the test to find out.
Let's check it out. So the same thing I did with GroqBot, I'm also trying with Manas. So I just asked Manas, can you just test my TestToolMatcher program? And I gave it the URL and let's see what happens.
Okay, so what's happening here is Manas is just starting to evaluate the page structure Without any prior context. So notice what I didn't do here. You know, I didn't give it page object model, a sitemap, or even a Swagger doc. I just pointed it to the URL and it automatically, autonomously spidered the DOM and figured out the intended user journey on its own.
And so as you could see, it reviewed the page structure, understands the tech stack. Now it's going to evaluate and try different test scenarios. So now it's going to run multiple test scenarios through the matcher, trying different things, capabilities, tech stacks, and user personas to evaluate how well this did. And once again, this is just Manus right out of the box.
I gave it nothing else.
And so let's see how it handles it. So it's just going through, navigating the application, exploring it. Running a test that it sees fit.
All right, so you see Manas hits the 404 page for the View All Tools and Details button. So right there, I just found an immediate routing issue for a zero configuration agent, just exploring the site. Finding a hard 404 on a primary CTA is really a massive win for the basic smoke test that anyone wants to create.
Now it's going into these different options for protocols when you select performance testing. So like I said, it knows nothing. It's exploring and it found all the different options based on what you selected in the first choice. Checking out the search box right now, seeing what's found when you type in it. Anything returns, right?
Cool.
So just typed in a random tool and it found it.
Valid page returned.
Let's see what happens. It's seeing if it's supposedly checking the quality of the output.
All right. So this is where the agent starts showing some real, I hate to use this word, but ROI. No, it's not just clicking buttons. It's actually translating what it found into a properly formatted actionable Create report complete with reproducible steps. This is exactly what a developer needs to fix the issue. Right, so no, I am being wicked lazy here on purpose. I wanted to see if we could create a closed-loop AI system. So I'm taking the exact bug report Manus generated and feeding it straight into Cursor as a prompt To see if one AI can fix the other AI's findings.
I'm going to take this Cursor prompt, I'm going to copy it, and then I'm going to just paste it right into Cursor and click run and see what happens. All right, so I'm in Cursor now. I'm opening up my project. I always say don't break anything. So before you run it, make sure you verify you don't break anything and then say, can you run this prompt that I was just given? All right, so it's running through. All right, so I checked in the code in GitHub and went back to Manus, said, okay, these bugs were fixed. Can you verify that these bugs are actually fixed now? And so Manus is going to go, okay, let me check it out.
All right, so Cursor pushed the fix, as you could see, and now Manus is doing the regression test. It successfully parsed my prompt, remembered the context of the previous bugs, and it actually validated the new deployment without me writing a single line of automation code.
And as you could see, it's saying it's retesting bug 1, bug 2. And you can watch its progress within the computer view as well, right? So it's saying it tested all the different examples and it confirmed whether or not they passed or failed, right? And then it gives a breakdown of all the bugs were actually fixed that it found previously. So this is an end-to-end testing lifecycle. Using just AI.
So this workflow really, I think, is incredible, but it exposes a massive risk if we just let the AI grade its own homework. So moving forward, a mature test chain would require probably multiple multi-agent systems where the AI writing the code is completely isolated from the AI testing the code. That tends to be a best practice. So I really would love to know, would you actually trust an AI agent to verify its own code fixes, or do you think a human tester is still essential? Let me know in my 11th annual State of Test Automation survey that I have in the link down below. Look forward to seeing your answers. And as always, test everything and keep the good.
Cheers.
