Harness provisioning follows a domain-dependent Pareto frontier, not a universal optimum

arXiv·medium signalXBlueskyLinkedInCopy link

Most systems hand every task the same comprehensive harness of tools, context and permissions. This paper treats harness selection as resource-matching, builds task-to-harness maps from both literature mining and controlled agent execution, and proposes map-guided escalation: start with a task-specific harness and expand to full provision only after a failed self-check. In liquid cooling operations, accuracy rises from 0.652 under full provision to 0.715 and matches Reflexion with 48% fewer tokens; in power grids full provision stays accuracy-optimal and the map only offers cheaper alternatives. The conclusion for builders is that trimming tool grants is a per-domain tuning exercise with a measurable accuracy cost, not a free optimization.