Blog.
Notes from our research into agents that learn through experience.
Ideas, methods, and experiments.
Learning Across Games: Why We Chose Batches
What our ARC-AGI-3 experiments taught us about shared skills, regression checks, and specialization.
Learning Facts, Learning Methods, Learning to Learn
What changes when an agent learns a task, improves its methods, and revises how it learns.
Recursive Language Models, Explained
How external context, code, and focused model calls make large tasks easier to work with.
Agents Shouldn’t Start Over
Why our work on a self-learning browser led us to study agents that improve through experience.