Morning Edition · Tuesday, August 4, 2026Published at 2:03 AM EDT · New York

New Benchmarks Target the Decision Agents Make Before They Answer

Two arXiv releases score the meta-decision of whether to answer, decompose, retrieve, execute code or delegate. A third replaces static persona prompts with synthesized lifelong memory.

Agentic systems spend most of their token budget on choices that precede the answer. MetaRoute-Bench formalizes exactly those choices, evaluating whether a controller should answer directly, decompose a task, call a tool, run code, delegate…