Found first: a primary source the press has not covered yet.
A paper posted to arXiv on August 27, 2026 introduces PILOT, a supervisor-worker harness for long-horizon agents that performs self-improvement during a run, not after it. Most self-improvement work operates on completed episodes; PILOT intervenes on active workers and converts observed runtime failures into reusable skills and memory in real time.
What the source says
The system, from Yang Xiao, Yusong Sun, Haoyi Wu, and seven co-authors, has two core mechanisms: live steering, which lets the supervisor redirect or terminate active workers mid-execution, and live self-evolution, which distills runtime procedures and failures into persistent skills and memory. Tested across two frozen model backbones and three benchmarks, PILOT ranked first in 5 of 6 configurations. On Terminal-Bench 2.0 it outperformed comparable systems by up to 9.8 percentage points. In the self-improvement setting, it gained 14.6 points with a GLM-5.1 backbone and 12.4 points with Kimi-K2.6. Mean output tokens fell 42.9% and 47.4% respectively, and successful evaluations per million tokens increased by 110.3% and 134.0%.
Why it matters
The architecture shifts agent self-improvement from a batch process run between episodes to something that happens inside a single run. Every execution becomes a simultaneous training episode, with the supervisor harvesting skills from failures that are still unfolding. The token reduction numbers matter separately: cutting output tokens 43-47% while improving task success rate means the efficiency gains are not a quality trade-off. For teams building long-horizon agents, the practical question raised here is whether a frozen backbone plus a capable supervisor can substitute for fine-tuning on task-specific data.