PILOT framework enables live self-improvement for long-horizon agents

rohanpaul_ai · x · 2026-08-31

Traditional agent self-improvement typically happens only after a task run concludes. The new PILOT framework introduces in-run learning by separating roles: a worker executes tasks while a supervisor monitors the trajectory, capable of intervening or aborting, and writes useful procedures or failure modes into persistent memory.

In experiments on Terminal-Bench 2.0 across 20 iterations, PILOT improved the best observed pass rate by 14.6 percentage points with GLM-5.1 and 12.4 with Kimi-K2.6, while reducing mean output tokens by 42.9% and 47.4% respectively. This mechanism allows the system to recover the current attempt, test immediately, and incorporate lessons without waiting for the next rollout.

Original post →

More from coding & agent

coding & agent channel →