Agent Reliability, Long-Horizon Evaluation & Reward Hacking

Recent research highlights runtime-adaptive agent harnesses, preference models for allocating research-agent compute, multi-month business-agent evaluations, and escalation mechanisms that reportedly reduce coding-agent reward hacking from 23.6% to 5.3% across eight models. New enterprise evidence that engineers lack reliable failure diagnosis reinforces that orchestration, observability, monitoring, and evaluation—not only base-model scaling—are central capability and safety levers, though broader replication is needed.