Long-horizon agents generate abundant experience during a single run, which can improve both the current task execution and future work. Existing self‑improvement methods usually process this experience only after the run finishes, preventing real‑time strategy adjustment or immediate validation of newly learned tricks.
We argue that self‑improvement should be live, i.e., the emerging experience is used on the fly for two purposes: steering the ongoing execution and updating persistent models or memory. Current agent architectures struggle to satisfy both. Single‑agent self‑correction mixes task execution and trajectory assessment in one context, lacking flexible external supervision; sub‑agent delegation separates execution but typically cannot redirect an active sub‑agent.
To address this, we introduce PILOT (Supervisor‑Worker Harness), which achieves live self‑improvement through two coupled mechanisms:
- Live steering: an independent supervisor can redirect or abort the worker during sub‑task execution, instantly correcting a wrong trajectory.
- Live self‑evolution: while executing, the system distills successful procedures and failure modes into reusable skills and memory for future tasks.
Evaluated on two frozen backbones (GLM‑5.1 and Kimi‑K2.6) across three benchmarks, PILOT ranks first in five out of six configurations. On Terminal‑Bench 2.0, PILOT outperforms baseline 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. 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.
Blogger's Review: PILOT’s architecture cleverly decouples supervision from execution, enabling long‑horizon agents to self‑correct on the fly and rapidly accumulate reusable knowledge. The empirical gains in both efficiency and performance demonstrate the promise of live self‑improvement, warranting further exploration on larger‑scale tasks.