Dari Is Betting Browser Agents Get Reliable by Using Less AI

There is a moment every builder of browser automations knows. The demo works. The agent glides through a login, fills a form, clicks submit, and everyone in the room nods. Then you run it a hundred more times and it starts clicking the wrong thing. Dari, a two-person company in Y Combinator's Fall 2025 batch, is built around that exact moment - and its answer is contrarian: use less artificial intelligence, not more.

Dari sells an API for building reliable browser-use agents. A browser-use agent is software that drives a real web browser the way a person would: reading the page, clicking buttons, typing into fields, moving through a multi-step task. The category exploded in popularity because it promised to automate the long tail of web work that never had a clean API - the internal dashboards, the government portals, the supplier sites that were built for human eyes only.

The problem is that most of these agents are non-deterministic by construction. They call a large language model on every run to decide what to do next. That makes them flexible, but also slow, expensive, and prone to random failure. Run the same task twice and you can get two different paths - one that works and one that quietly breaks. For a demo, that's fine. For a business process that has to run a thousand times a day, it's a liability.

Consider what "unreliable" costs in practice. A finance team that automates invoice entry cannot tolerate an agent that succeeds 95 percent of the time, because the missing five percent is not evenly distributed - it clusters on the exact edge cases a human would have caught. A support team pulling data from a vendor portal cannot have the job silently return yesterday's numbers because the agent wandered down a different path. The whole value of automation is that you stop watching it. An agent you have to babysit is not saving anyone time; it is just relocating the work.

01 / The IdeaLearn once, then stop guessing

Dari's core move is to separate the thinking from the doing. On the first run of a workflow, it uses an LLM to figure out the path through a site - which elements to click, what to type, where the task goes next. Then it caches those steps as deterministic DOM instructions. Every run after that replays the cached path directly, without paying for another model call. The intelligence gets spent once, up front, and the repetitive part becomes cheap and predictable.

The last step is where the design earns its keep. Websites change - a button moves, a field gets renamed, a layout is redrawn. A brittle script would simply fail. Dari treats a change as a signal: it detects the mismatch, calls AI again to find the new path, and caches that. The failure teaches the system something instead of just breaking it.

It is a small inversion of the usual instinct, and it is the whole company. The prevailing approach in the agent world has been to make the model smarter and call it more often, on the theory that intelligence solves reliability. Dari's wager is the reverse: that reliability comes from consulting the model rarely and remembering aggressively. Most of the web is repetitive. The pages you automate today look like the pages you automated yesterday. Spending a fresh round of reasoning on an unchanged form is not thoroughness - it is waste, and waste at scale is where the bills and the flakiness both come from.

02 / The Hard PartWhere automations go to die: the login

Ask anyone who has tried to automate real web work and they will point to the same wall: authentication. The workflows worth automating are almost always behind a login, and modern logins mean two-factor authentication. A one-time code arrives by text message, or lives in an authenticator app, or lands in an inbox - and most agents simply cannot get past it.

Dari's answer is to handle 2FA directly. It manages SMS text codes, TOTP authenticator codes, and email verification automatically, so a workflow can log in, get past the challenge, and keep going without a human waiting to paste a six-digit number. That single capability is what turns browser automation from a party trick into something that can touch the systems a business actually runs on.

It sounds like a footnote until you notice how much of the useful web lives behind it. The interesting data is rarely on the public marketing page; it is inside the account, past the sign-in, guarded by a code. An automation tool that stops at the login can only reach the parts of the internet that were already easy. By walking through the door most agents get stuck at, Dari changes which tasks are even on the table - not the demos, but the authenticated, unglamorous back-office work that is expensive precisely because it has always needed a person.

03 / The ShapeAn automation you can call like an API

Dari does not stop at making the browser part reliable; it wraps the whole thing in developer plumbing. An agent project becomes a versioned, stateful API endpoint. You can describe an automation in plain English to author it, then trigger it with a POST request or a webhook. Publishing a new version keeps a stable URL, and long-running sessions are designed to survive crashes and resume where they left off. The result is that a messy, human-only web task starts to behave like any other backend service a team can call, monitor, and version.

That framing - reliable, versioned, callable - is what separates Dari from the two things it competes with. On one side are the general browser agents that dazzle in a demo and wobble in production. On the other are traditional scripting stacks like Playwright and Selenium, which are deterministic but brittle: they break the moment a site changes and have no idea how to recover. Dari is trying to sit in the gap, keeping the determinism of a script and borrowing intelligence only when the script would have failed.

04 / The PeopleTwo engineers who hit the wall themselves

Dari was founded in 2025 by Avyay Varadarajan and Benjamin Hong, who met studying computer science at Caltech and went on to engineering roles at companies including Amazon and Uber. Varadarajan, the CEO, has also spent time on machine-learning work outside of pure product engineering. The company is based in San Francisco and, as of its launch, is still just the two of them - a detail that fits the product, which is essentially the fix for a problem they kept hitting while building with browser agents.

That origin matters for how to read the company. Dari is not a grand thesis about the future of autonomous software. It is a narrower, sturdier claim: that a large slice of web work is the same steps repeated, and that the smart thing to do with those steps is remember them rather than re-derive them. The pitch is almost boring, which is part of why it is credible.

A small team also has a way of sharpening the product. With two people and no room for a sprawling roadmap, the decisions show up plainly in what the tool does: cache the steps, handle the login, expose an endpoint, adapt when a page moves. There is no feature built to impress a slide. Everything traces back to the same lived complaint - that the agents they wanted to rely on could not be relied upon - and that focus is easier to keep when the company is still small enough to fit in one conversation.

05 / The MarketInfrastructure, not magic

The browser-automation space is crowded with names promising agents that can do anything on the web. Dari is competing less on ambition and more on temperament. Its bet is that the winners in this category will be the ones whose automations run correctly the ten-thousandth time, quietly, without a person watching. That is an unglamorous place to plant a flag, and it is also where real businesses spend real money.

There is a broader pattern here worth naming. Early in any technology wave, attention goes to what is newly possible - the flashy first run, the thing that could not be done before. Money and durability tend to arrive later, when the same capability becomes something you can depend on. Databases got boring. Payments got boring. The web itself got boring, in the best sense. If browser agents follow that arc, the reliability layer is not a smaller prize than the intelligence layer; it is the part that survives the hype and keeps charging rent.

Whether Dari becomes the default layer for reliable browser agents or one option among several will depend on how well the caching-and-adapting loop holds up across thousands of messy, ever-changing sites. But the framing is clear and the problem is real. In a field full of demos, a company selling the dishes - the repetitive, authenticated, boring-but-critical work - has picked a fight worth having.