Training AI to revise agent harnesses using execution feedback

Researchers describe “harness learning,” which trains an AI model to revise an agent’s harness—the program organizing model calls, tool use and information flow—using feedback from task runs. They report better revisions in reasoning and multi-hop question-answering experiments, with the ability to adapt transferring to unseen tasks. Revisions could keep improving harnesses over multiple rounds, though training on revision sequences had varying benefits.