The new Exo framework allows AI models to boost their task performance by iteratively modifying their own underlying architecture while executing work. While agent harnesses are typically hand-tuned or optimized against benchmarks before deployment, Exo provides models with runtime access to logs and the ability to rebuild their own structures, a process the developers say has yielded surprisingly strong results.
To prevent the autonomous editing process from compromising the agent’s stability, the system incorporates safety protocols adapted from operating systems research. The project is framed as a form of online continual learning, designed to let agents log successful configurations and evolve a general skill for adapting their operational harnesses to new tools and environments.
