PeakBench separates an agent's dependency planning from its resource scheduling, and finds good planning does not produce safe parallel execution

arXiv·medium signalXBlueskyLinkedInCopy link

PeakBench (arXiv 2608.24509, 25 August) benchmarks something existing agent benchmarks skip: whether an agent can parallelize tool calls safely under resource constraints, rather than just selecting the right tool and arguments under mostly serial execution. It provides executable multi-tool workflows with execution-grounded dependency annotations and measured resource profiles, plus a two-part evaluation that disentangles logical planning failures from physical scheduling failures so a bad run can be attributed to one or the other. The headline result is that strong logical planning does not reliably translate into safe or efficient execution under resource limits, and that simply exposing resource information to the agent reduces avoidable overflows. Code is at github.com/Czzzk/Staggering-the-Peaks.