What happened
Researchers built a system where cheap, fast AI models can ask stronger (but more expensive) models for help only when they actually need it, instead of always using one or the other. This lets companies run AI agents that solve complex problems without paying for top-tier AI on every single step.
Why it matters
Right now, companies choose between slow-but-reliable AI or fast-but-mistake-prone AI for handling multi-step tasks. This shows you could use cheap AI for 80% of the work and only escalate to expensive AI when the cheap version signals it's stuck — if it works at real scale, that's a way to cut AI inference costs without sacrificing accuracy on hard problems.