We built Smooth because existing browser agents were frustrating. They were slow, expensive, and unreliable. Even simple tasks could take minutes and cost dollars in API credits.

We started as users of Browser Use, but the pain was obvious. So we built something better. Smooth is 5x faster, 7x cheaper, and more reliable. And along the way, we discovered two principles that make agents actually work.

(1) Think like the LLM [2]

The most important thing is to put yourself in the shoes of the LLM. This is important when designing tools, but even more important when designing the context. How you present the problem to the LLM determines whether it succeeds or fails. Imagine playing chess with an LLM. You could represent the board in countless ways - image, markdown, JSON, etc. Which one you choose matters more than any other part of the system. Clean, intuitive context is everything. We call this LLM-Ex.

(2) Let them write code [3]

Tool calling is limited. If you want agents that can handle complex logic and manipulate objects reliably, you need code. Coding offers a richer, more composable action space. Suddenly, designing for the agent feels more like designing for a human developer, which makes everything simpler. By applying these two principles religiously, we realized you don't need huge models to get reliable results. Small, efficient models can get you higher reliability while also getting human-speed navigation and a huge cost reduction.

Here's how it works: 1. Extract: we look at the webpage and extract all relevant elements by looking at the rendered page. 2. Filter and Clean: then, we use some simple heuristics to clean up the webpage. If an element is not interactive, e.g. because a banner is covering it, we remove it. 3. Recursively separate sections: we use several heuristics to represent the webpage in a way that is both LLM-friendly and as similar as possible to how humans see it.

We packaged Smooth in a super easy-to-use API with instant browser spin-up, custom proxies, persistent sessions, and auto-CAPTCHA solvers. Our goal is to give you this infrastructure so that you can focus on what's important: building great apps for your users.

Before we built this, Antonio was at Amazon, Luca was finishing a PhD at Oxford, and we've been obsessed with reliable AI agents for years. Now we know: if you want agents to work reliably, focus on the context.

Try it for free at https://zero.circlemind.co/developer

Docs are here https://docs.circlemind.co

Demo video https://youtu.be/18v65oORixQ

We'd love feedback :)

[1] https://docs.circlemind.co/performance

[2] https://x.com/karpathy/status/1937902205765607626

[3] https://arxiv.org/pdf/2401.07339