Attributing utility to individual trajectory steps by counterfactual replay, so you know which agent steps you are paying for

arXiv 2608.26195·low signalXBlueskyLinkedInCopy link

arXiv 2608.26195 (2026-08-25) frames the core measurement problem in agent economics: cost accrues per step in tokens, latency, money and risk, while utility only appears at the end of the task. The proposed loop is cost profiling, then utility attribution, then misalignment diagnosis, targeted adaptation and evaluation, with attribution methods ranging from cheap process proxies up to counterfactual replay. It catalogs five misalignment shapes across reasoning, external calls, recovery loops, resource allocation and multi-agent coordination; it is a framework paper with no benchmark numbers, so treat it as vocabulary for instrumenting your own runs.