Skills Are the New Agents

Hard Problem in Agentic AI Is No Longer Reasoning, but Selection

Large language models are increasingly asked to behave less like passive text generators and more like autonomous problem solvers. In response, the AI community has embraced multi agent systems, where multiple specialized agents collaborate through explicit communication. This architecture works, but it is expensive. Every agent call repeats context, burns tokens, and adds latency.

A recent study, When Single Agent with Skills Replace Multi Agent Systems and When They Fail, reframes the problem in a sharper way. Instead of asking how many agents we need, it asks a more fundamental question: can we compress a society of agents into a single model that chooses among a library of skills, and if so, where does this approach break down?

The answer turns out to be both promising and unsettling. Skill based agents can match multi agent performance at a fraction of the cost, but only up to a point. Beyond that point, the system does not degrade gracefully.