Exo aims to let AI agents edit their harnesses while protecting state and secrets

Arena shared a presentation by Google Research senior researcher and UC Berkeley PhD candidate Alex Krentsel on Exo, an architecture designed to let AI agents edit their harnesses at runtime. Borrowing an idea from operating systems, Exo separates protected state, history, secrets and a “guardian” from an evolvable policy layer. The presentation says Exo was 31–62% cheaper than Codex and Pi on open-world RuneScape tasks.