Exo: An Architecture for Self-Improving Systems at Runtime
Alexander Krentsel, Sylvia Ratnasamy, Ion Stoica (UC Berkeley), Martin Casado (Andreessen Horowitz) and Ankur Goyal (Braintrust) present Exo, an agent runtime in which the agent revises its own harness code during ongoing work instead of in a separate offline optimization loop.
Ask this paper
Design. A small protected mechanism layer keeps execution history and controls which policy version is active. Everything above it, the harness policy, can be changed by the agent through inspectable code edits.
Open-world game. On long-running RuneScape objectives (combat, crafting, currency), Exo completes tasks at 12 to 53% lower model cost than Codex and Pi in standalone runs and 45 to 82% lower in transfer runs, at matching completion rates.
Held-out tasks. Against the same starting policy without self-editing, Exo cuts held-out model cost by 35 to 61% on AppWorld and SREGym with comparable correctness, reducing AppWorld mean task time by 13% and SREGym median diagnosis time by 35%.
Coding. On Continual Learning Bench coding tasks, total model cost including learning falls by 14%, and both policies resolve every issue.
Abstract
Agent harnesses govern how language models act in an environment; designing their policies by hand is difficult, while automated techniques require representative datasets and costly executions. We present Exo, an agent runtime architecture for safe learning during ongoing work that collapses the optimization loop into task execution, allowing agents to revise both their harness and their strategies for further improvement. A minimal protected mechanism layer preserves execution history and governs policy activation while the harness evolves through inspectable code changes. On long-running objectives in an open-world game, Exo reduces model cost by 12-53% relative to Codex and Pi in standalone runs and 45-82% in transfer runs, while matching task completion rates. Compared with the same starting policy without the ability to edit policy, Exo reduces held-out model cost by 35-61% on AppWorld and SREGym with comparable correctness, cutting AppWorld's mean task time by 13% and SREGym's median diagnosis time by 35%. On Continual Learning Bench coding tasks, model cost including learning falls by 14%, with both policies resolving every issue.