This paper introduces PILOT, a supervisor-worker agent harness that performs self-improvement live during a run rather than only after it ends. A separate supervisor can steer or abort an active worker mid-execution, while a second mechanism distills procedures and failure modes into reusable skills and memory in real time. Across two model backbones and three benchmarks, PILOT ranks first in five of six configurations and outperforms other harnesses by up to 9.8 points on Terminal-Bench 2.0, while cutting output tokens by over 40%.
