PILOT harness steers agents mid-run, cuts output tokens 42.9%

The core claim of the new PILOT paper on arXiv is procedural, not architectural: a separate supervisor watches an agent run and yanks it around while it is still in motion.

"Live steering lets a separate supervisor redirect or abort the active worker during execution," the abstract states, paired with a second mechanism the authors call "live self-evolution" that "distils procedures and failure modes revealed during execution into reusable skills and memory." Prior self-improvement work, they argue, only mines a trajectory once it is done. PILOT mines it while it is still running.

The headline numbers. On Terminal-Bench 2.0, PILOT beats counterpart harnesses by up to 9.8 percentage points, and across two frozen backbones and three benchmarks it ranks first in five of six configurations. In the self-improvement setting Yang Xiao and colleagues report gains of 14.6 points with GLM-5.1 and 12.4 points with Kimi-K2.6.

The efficiency line is where the paper gets loud. Mean output tokens fall by 42.9% and 47.4% on those two backbones. Successful evaluations per million output tokens rise 110.3% and 134.0%. Roughly, each dollar of inference buys about twice as many completed tasks.

The abstract does not name the counterpart harnesses, list the other two benchmarks beyond Terminal-Bench 2.0, or quantify the supervisor's own token cost. The one configuration where PILOT does not place first also goes unnamed. Results are on two frozen backbones only.