Our thesis
"The best way to predict the future is to invent it."
Pearl started as something we thought about in our grade 9th, one of the coolest things we ever watched at the time (it still is) was Made in Abyssand we were super intrigued by a character named Reg who was a robot yet had human emotions, felt the same pain and ate human food, it acted as a guide to the protagonist to fulfill her journey.
Inspired by this, we wanted to create something similarly human-like, but for a post-AGI world, where many agents and assistants had become mere tools—lacking personality and a humane touch. We previously tried building Asta, an agent to handle day-to-day tasks, and our complaint was the same: no real advancements over the current ecosystem.
Shreyansh and I always dreamed of something where we could sit and talk about our projects, ask it to edit our sheets, Todoist, or documents, and navigate through the project together, deciding on things in real-time. We wanted a true assistant—one that doesn't just talk, but also takes action.
Imagine a friend who's available not just to quickly relay and reply to messages on your behalf, but also to sit with you for half an hour, talk through a proper roadmap for your project, set up your linear for it, create a planning document, and then tweet it out for you. That's exactly what we wanted to create.
Pearl, in order to perform tasks with minimal prompt layering, has full context awareness across the apps you give it access to. We don't use screen recorders or active watchers; everything happens during execution, preserving privacy. In the future, we'd love to explore local LLMs to bundle up with applications.
Pearl is essentially like a conscious AI that takes actions just like we humans do. It asks questions when you give it tasks—asking why you want it, how you want it, and when you want it. Pearl is essentially inspired by Reg from Made in Abyss, a character we admire.
How do we do it? It's nothing special. We started with custom function calling during Asta, essentially a smart prompt-to-function caller. Then, during the launch of Rabbit R1, the initial failure and limitations of function calling encouraged us to try browser agents. We built a custom Chromium fork, but TLDR: it didn't work out. There were many reasons: it wasn't safe, it wasn't reliable, and it felt like bringing a nuke to a sword fight—not practical.
During our research, and thanks to advancements in LLMs, we narrowed down an approach that was the best of both worlds. It's nothing fancy, but it's a simple yet very effective solution (we can't reveal it yet). Based on our data, vanilla models currently perform well over 80% of tasks across 15 widely used applications, and we aim to bump that up to 95% by Q1 2025.
Pearl also has built-in improvisation—whenever it fails, it self-heals, fixes the problem, and completes the task. We believe that humans take around 5-10 minutes to complete certain tasks, while Pearl solves them in under 60 seconds. We prioritize accuracy and human-like results over speed.
All in all, with Pearl, we want to build a real interface, not just an agent—something that feels alive, beyond a 16:9 chat window on your screen.