Introduction
Today, we’re excited to announce Dojo, a collaborative RL environment suite for computer use agents (CUA).
Dojos are interactive CUA environments where models learn by doing.
They can interact with elements on screen, play games, solve puzzles, perform actions on productivity software - and learn from rewards defined from these tasks.
Dojo allows you to seamlessly spin up interactive environments with tasks - allowing you to train and evaluate top CUA models against like Sonnet 4.5, GPT-5/4o, and self-hosted models.
Motivations
Historically, frontier labs have worked with data vendors to build clones of top websites internally.
These clones are useful because tasks can be written against them without having to worry about changes in UI, CAPTCHA, or authentication.
Dojo brings these clones from the closed off walls of frontier labs to the general public, so anyone can train on them and contribute towards them.
Examples
Ex1: Agent organizing Linear tickets (on a Dojo clone)
Ex2: Agent searching for leads on LinkedIn (on a Dojo clone)
Ex3: Agent playing (and intentionally losing) Tic-Tac-Toe
Integrations
In addition to out of the box support for evaluation, Dojo has first class support with popular RL frameworks like:
