The dream was simple: tell your AI agent to book your flights, compare CVs, or fill out your taxes while you relax. Instead, you’re solving CAPTCHAs for a robot that’s supposed to be solving problems for you.

We tried out Manus. Leave it to Manus, they claim. It’s one of the buzziest new "autonomous agents" in the AI startup world, and while promising, it spent more time being confused by Cloudflare than executing my tasks. The experience offered a strange déjà vu: instead of streamlining workflows, these agents are now trapped in an arms race with the very fabric of the modern internet.

CAPTCHA is the new firewall

AI agents operate by browsing the web like humans do, only they’re not humans, so the web increasingly treats them like malicious intruders. CAPTCHA, once just a mild annoyance, has become the default border control of the internet. The results? Your AI agent hits a wall, throws up its hands, and asks you to do the busywork.

Ironically, these bots are falling victim to the same dark UX patterns humans have had to suffer for years. And even more ironically, these patterns were used to train and create AI models. We’ve built a web that’s actively hostile to efficiency. Popups, login walls, obfuscated buttons, and paywalls that shift like sand.

It’s annoying. It's structural. It's by design.

The internet wasn’t broken. Then we monetised it.

Websites weren’t always adversarial. But the modern internet has warped itself around monetisation models that rely on user friction. SaaS platforms increasingly gate features behind subscription paywalls, inject dark patterns to nudge upgrades, and structure their UX to maximise profit over usability.

It’s not just a bad user experience. It’s bad for automation, too. AI agents are exposing this by simply trying to get stuff done. But when “getting stuff done” means crawling through deceptive modals, rotating IPs, or pretending to be Firefox, you realise the system isn’t broken, it’s working exactly as intended.

AI won’t save you from bad UX

Some founders see AI agents as a UX band-aid: why fix the interface when you can just build an agent that clicks all the buttons for you? But that only adds a layer of technical debt on top of your existing design debt.

Automating the pain doesn't erase it, it multiplies it. If your application is designed to be inconvenient in order to extract value from your users, AI will simply experience that inconvenience faster and louder. And maybe blog about it.

Pay-to-crawl: the future of data access

It’s only a matter of time before web data access becomes a subscription feature. If agents want to get through the front door without CAPTCHA hurdles, they’ll need to pay rent.

We’re already seeing this shift. Some content platforms are considering metered APIs, agent-specific access tiers, or embedding attribution requirements. AI agents will become SaaS users themselves. And that’s probably a good thing.

Open source: the last line of defence

While commercial platforms are busy obfuscating access, the open source world offers a radically different ethos. Open source products like Jellyfin or Ghost don’t have incentives to lock you out. They thrive when you thrive.

Of course, open source has its own problems: support, visibility, and feature velocity. But it doesn’t try to trick you. It doesn’t invent pain to sell you the cure. And that’s worth celebrating.

Final thoughts: don't build like the worst parts of the web

As founders and investors, it’s worth asking: are we building value, or are we building friction? Are we optimising for delighted users, or just for better funnel metrics?

The web doesn't need more manipulative interfaces and $9.99 feature unlocks. It needs honest, usable, sustainable products. AI agents may not rescue us from the internet’s enshittification — but perhaps they can at least hold up a mirror.

And if they do get stuck in a Cloudflare CAPTCHA loop, maybe that’s your cue to fix the real problem.

Curious how we think about engineering and automation? At madewithlove, we help SaaS companies build digital products — the kind that don’t require a bot to use them. See how we help product and engineering teams.