"Beyond Component Testing": agentic AI needs trajectory validation, not component validation

arXiv·medium signalXBlueskyLinkedInCopy link

A July 31 position paper (arXiv 2607.29405) argues that trustworthy agent deployment depends on validating trajectories in context rather than assessing isolated components, because agent behavior emerges from multi-step decision sequences shaped by planning, tool use, memory and a changing environment. It organizes validation across five dimensions — behavioral, safety, temporal, regulatory, and multi-agent — and identifies temporal validity and runtime evidence maintenance as the biggest gaps: a system validated in March is not validated in August if the environment moved. The proposed agenda is bounded-autonomy specifications, adversarial trajectory testing, continuous monitoring, and audit-ready documentation. It pairs with the ProofAgent Index work from earlier this week; both are converging on the claim that capability benchmarks do not predict production readiness.