Someone posted on Hacker News last week about building an AI browser agent to automate Reddit because the official API was either paywalled or nonexistent for most use cases. The comments split predictably: half calling it a hack, half asking for the repo link. But the interesting part is not the Reddit project itself. It’s that this exact pattern keeps showing up in completely unrelated contexts, and I think it tells us something real about where integration is heading.

The API gap is bigger than you’d guess

We’ve all been trained to assume any major platform has an API for anything useful. But this hasn’t been true for a while. LinkedIn gutted its public API years ago. Twitter killed free API access entirely and priced enterprise at $42,000 a month. And most government portals expose zero programmatic access. Your health insurance provider, your state’s DMV appointment system, your company’s internal HR tool running on some vendor platform from 2014: none of these have APIs, and none of them ever will, because the organizations behind them have absolutely no reason to build one. The assumption that every web service comes with an API is a developer bubble. Most software in the world was built for humans to click through, and it stops there.

And even when APIs do exist, they frequently don’t cover the thing you actually need. Reddit’s API is real but does not support bulk moderation the way the web UI does. Salesforce has one of the most extensive APIs in enterprise software and people still build browser automation around it because some admin workflows require clicking through four nested modal dialogs that have no programmatic equivalent whatsoever. So browser agents fill this gap. Not just where an API is too expensive, but where no API was ever going to exist in the first place.

Reddit was the canary

The HN post got traction because Reddit has a well-known API story. The 2023 pricing changes killed most third-party clients, and Reddit’s API situation has been a sore point for developers for three years. But Reddit is actually one of the easier cases. Or at least it has a partial API you can cobble something together with.

The harder cases never make Hacker News because they’re too boring. Filling out procurement forms in a company vendor portal. Checking shipment status across three logistics platforms that all require manual login. Pulling data from government databases that render results as paginated HTML with no export option. Submitting compliance reports through portals built by the lowest bidder in 2009.

And browser agents handle these because nothing else can.

What browser agents actually replace

Concrete examples of tasks people run through browser agents because no API exists:

Comparing prescription drug prices across pharmacy websites. GoodRx has an API; your local pharmacy chain does not. You open five tabs and squint. Or you let an agent do it.

Monitoring state government contract postings. A few states publish RSS feeds. But most just slap up a search page you need to check by hand every morning.

Submitting expense reports through corporate finance tools built by vendors whose “partner API” costs six figures a year. So you get the web form, and you click through it every Friday.

Booking government appointments. A web portal exists. An API does not and will not.

And I wrote about more of these daily browser tasks in 7 Tasks You’re Still Doing Manually in Your Browser.

Why browser agents work now and didn’t five years ago

RPA tools tried this approach for a decade. UiPath and Automation Anywhere sold expensive licenses to enterprises for brittle scripts that shattered whenever someone moved a button three pixels. And robotic process automation became a punchline in most engineering orgs I’ve talked to.

But LLM-powered browser agents adapt when pages change, because they parse meaning rather than memorizing CSS selectors and pixel coordinates, which turns out to be a massive difference in practice even if it sounds incremental on paper. dassi uses this approach in Chrome’s side panel, working on whatever page you’re looking at, and I keep hearing from people who automated some terrible vendor portal within ten minutes of installing it. And the gap between “theoretically possible with RPA” and “actually works reliably with an LLM” is where this whole category lives now.

The accidental guarantee

Organizations that refuse to build APIs have accidentally guaranteed browser agents will eat their lunch. They built a web portal for humans, and then an AI learned to use it. Nobody asked permission.

So where does this leave us

Every SaaS tool, internal dashboard, and government form is already accessible through a browser, which means a single browser agent can replace hundreds of API integrations that would each need separate clients, separate auth flows, and separate ongoing maintenance if you tried the traditional approach. And I think browser agents become the default integration method for any system without an API, which is most systems. That is the part people keep glossing over in conversations about AI automation.

If you want to see what this feels like, dassi is free. Bring your own LLM key, no subscription. Damn near any website works.