AI agents can burn 30 times more tokens for the same task

A new paper from researchers at Stanford, Michigan and other institutions finds that AI agents can use up to 30 times more tokens to finish the same task. The extra spending often buys no better result. In a field where compute is the scarce, expensive input, that is an awkward finding. Much of the cost of running agents may be wasted effort rather than needed thinking, which is why routing and efficiency tools are suddenly in demand. The reported figure is the upper end of the range, not the typical case.