Exo Architecture Enables Safe Runtime Self-Improvement for AI Agents
UC Berkeley and Google Research introduced Exo, an architecture inspired by exokernels that separates AI agents into a protected mechanism layer and an evolvable policy layer. This design enables safe runtime self-improvement of agent harnesses, demonstrating 31-62% cost reductions compared to Codex and Pi in benchmarks and open-world tasks.
Full Report
Arena.ai (@arena) · 2026-10-29 17
What if AI agents could safely improve their own harness at runtime? UC Berkeley PhD candidate and senior researcher @AlexKrentsel from @GoogleResearch recently presented his work on self-improving AI agents at Arena HQ.
This architecture, Exo, borrows the exokernel idea from operating systems, separating an agent into a protected mechanism layer (state, history, secrets, and a "guardian") and a fully evolvable policy layer.
Alex walks through the design, results on benchmarks and open-world RuneScape tasks, where Exo was 31-62% cheaper than Codex and Pi, and more.
Big thanks to @AlexKrentsel for joining us to present his work with @istoica05, @martin_casado, @ankrgyl, Sylvia Ratnasamy, @BerkeleySky, and @netsyslab.
