Simple web-agent routing cuts cost up to 30.6%

A simple fallback to the cheapest observation mode cut web-agent cost 9.5% to 30.6% across all eight tested site-model combinations without lowering task success. The new routing preprint compared six text-and-vision modes and found that rerunning the same setup changed 12% to 14% of outcomes, while five learned policies generally failed to beat one well-chosen fixed mode. Teams following production-grade evaluation discipline for autonomous agents should measure repeat-run variance and test a cheap failure cascade before funding a learned router whose apparent upside may be noise.