TinyFish

The gap is not intelligence. It is plumbing. A large language model is trained on text; a website is a live, twitchy, login-gated interface that changes its buttons on a Tuesday for no reason. Ask a general-purpose agent to log in, page through inventory, fill a form, and come back with clean data, and it will often do the first three steps beautifully and then face-plant on step four, forever, in a way that is expensive and hard to notice. TinyFish's founders looked at this and made an argument that is either obvious or heretical depending on your mood: roughly 92% of the useful web sits behind logins and dynamic interfaces that today's agents cannot reliably reach, and somebody needs to build the boring infrastructure layer that lets them.

That is the whole company. TinyFish, founded in 2024 and launched publicly in August 2025, builds what it calls enterprise web agents - software that navigates web applications with something like human judgment but at machine speed and scale. The pitch is not "we automate tasks." The pitch, in the words of chief executive Sudheesh Nair, is that "helping companies get more value from the web isn't about automating low-value tasks. It's about amplifying the high-value, outcome-driven processes that require human-like interaction at scale." Which is a very enterprise-software way of saying: we do the web chores that are too fiddly for a script and too dull for a human, and we do a lot of them at once.

The most quotable example is Google. Google Travel would like to show you hotels in Japan, including the small properties running booking systems that predate the smartphone and expose no API whatsoever. There is no integration to build because there is nothing to integrate with. So TinyFish's agents simply use the websites - the way a very fast, very patient person with thousands of hands would - and surface that inventory into Google Travel. No infrastructure changes on the hotels' side. This is a deeply unglamorous problem, and the fact that a $2 trillion company outsources it is the entire investment thesis in one anecdote.

The economics of that are worth sitting with. The value of a web agent is not that it can do a thing once, in a demo, on a stage. Anyone can build that; the internet is littered with agent demos that work exactly one time. The value is that it keeps working next week, when the hotel's website moves a button or renames a field, without a human being paged at 2 a.m. TinyFish's patented infrastructure is designed to learn and adapt rather than shatter - agents that treat a changed page as a new situation to reason about, not a fatal exception. Reliability is the unsexy moat, and TinyFish has decided to make the entire company about it.

It helps that the founders are, by startup standards, adults. Nair was president of Nutanix and later chief executive of ThoughtSpot - a man who has sold expensive software to nervous enterprises and knows they buy outcomes, not vibes. Co-founder Shuhao Zhang built massive-scale systems at Meta, which is the relevant credential when your product is "do millions of web operations a month without falling over." Co-founder Keith Zhai came from the newsroom - a senior correspondent at The Wall Street Journal - and handles positioning and communications, which is a polite way of saying he is the reason the goldfish has a personality.

Investors noticed. In August 2025, TinyFish announced a $47 million Series A led by ICONIQ, with USVP, Mango Capital, MongoDB Ventures, ASG, and Sandberg Bernthal Venture Partners along for the ride. ICONIQ's Amit Agarwal offered the sort of quote investors offer - "TinyFish has the technology, leadership, and momentum to become one of those enduring companies" - but the interesting part is the shape of the bet. This is not a consumer app you will ever see. It is invisible plumbing that other companies' AI runs on top of. Writing a $47 million check for something no end user will ever notice is a specific kind of conviction, the kind you have when you think the invisible layer is the one that ends up mattering.

What can you actually do with it? If you are a large company, quite a bit of the drudgery you currently pay people or outsourcing firms to do by hand: monitoring competitors' prices across thousands of sites in real time, tracking inventory that lives only in someone else's web portal, pulling market intelligence, running insurance-quote workflows, doing social listening, or executing autonomous QA testing on your own products. The through-line is that all of these involve a computer patiently operating a website meant for a human, at a volume no human would tolerate. TinyFish reports running hundreds of thousands of agents performing millions of operations a month, which is either impressive or slightly unnerving, and probably both.

For developers, TinyFish also exposes the machinery as a set of APIs under one key - a Search API that returns structured JSON from dynamic pages, a Fetch API that turns rendered pages into clean markdown, an Agent API for multi-step automation with logins and forms, and a Browser API offering stealth Chromium sessions that cold-start in under 250 milliseconds. Search and Fetch are free; Agent and Browser are pay-per-credit. This is the classic land-and-expand move - let an engineer try the cheap tools on a weekend, then sell the enterprise version to their boss when the weekend project becomes a business dependency.

There is, lurking under all of this, a genuinely thorny question the industry has not resolved: what does it mean for the web when a growing share of its visitors are agents wearing stealth browsers to slip past bot detection? TinyFish frames its work around consent, compliance, and legitimate enterprise use, and its marquee customers are the sort who employ large legal teams. But the arms race between automated access and anti-bot defenses is real, and a company whose product is "reliably act on any website" is going to keep bumping into it. That tension is not a knock on TinyFish so much as the weather it operates in.

For now, the story is unusually clean for an AI startup: a specific, boring, valuable problem; customers who are already paying; founders who have done the enterprise thing before; and a mascot that looks like it belongs on a children's cereal box. In a market drowning in agents that can talk about the web, TinyFish is betting on the ones that can quietly, reliably, unspectacularly get the web to do what you asked. That is a smaller claim than most AI companies make. It might also be a more useful one.