Hugging Face researchers have unveiled PILOT, a novel supervisor-worker harness designed to enable live self-improvement for long-horizon AI agents. Unlike traditional methods that process execution experience only after a run ends, PILOT uses emerging experience to redirect the active run and update the persistent harness in real time.
The system couples two mechanisms: live steering, where a separate supervisor can redirect or abort the active worker mid-execution, and live self-evolution, which distills procedures and failure modes into reusable skills and memory. This addresses a gap in existing architectures, where single-agent self-correction mixes execution and assessment, and subagent delegation cannot redirect active subagents.
Across three benchmarks and two frozen backbones (GLM-5.1 and Kimi-K2.6), PILOT ranks first in five of six configurations. On Terminal-Bench 2.0, it outperforms counterpart harnesses by up to 9.8 percentage points. In the self-improvement setting, it gains 14.6 points with GLM-5.1 and 12.4 points with Kimi-K2.6.
Efficiency metrics are notable: mean output tokens drop by 42.9% and 47.4%, while successful evaluations per million output tokens rise by 110.3% and 134.0%, respectively. The authors plan to release code on GitHub soon.