Most agent self-improvement happens too late: the system learns from a run only after that run is already over.

This new paper, moves that learning inside the run with PILOT, the harness.

A worker handles the task while a separate supervisor watches its trajectory, can steer or abort it, and writes useful procedures or failure modes into persistent skills and memory.

That division lets the worker focus on execution while the supervisor keeps a cleaner view of the goal, progress, and drift.

Using the same frozen backbone for every system, PILOT ranked 1st in 5 of 6 model-benchmark configurations. Across 20 Terminal-Bench 2.0 self-improvement iterations, its best observed pass rate improved by 14.6 percentage points with GLM-5.1 and 12.4 with Kimi-K2.6. Mean output tokens per evaluated task fell 42.9% and 47.4%, respectively.

For long-horizon agents, that changes the self-improvement loop: a lesson does not have to wait for the next rollout. It can recover the current attempt, get tested immediately, and become part of the system used by later workers.

– arxiv. org/abs/2608.26530

Title: "PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents"