Researchers built a benchmark that puts AI agents through a version of the marshmallow test, and most of them cave.
The setup, described in a new arXiv paper, runs ReAct-style agents minute-by-minute with a tool that lets them "raise a question" instead of acting immediately, under a strict per-step budget on how often they can use it. The researchers then vary the conditions: agents work alone or in view of other agents (broadcast vs. isolated), get assigned personas with different ages and "hedonic drive," and are either required or merely allowed to use the self-questioning tool. Across 19,200 trajectories in 64 experimental cells, only 75.9% of agents made it to the end without an early "eat" impulse taking over. Isolated agents held out longer than ones broadcasting to peers, and agents forced to keep questioning gave in faster than those given a choice.
The interesting part isn't the marshmallow gimmick, it's the methodology. Most agent benchmarks are pass/fail snapshots; this one tracks behavior minute by minute using survival statistics, the kind of tool used for time-to-event data. That's a more rigorous way to study what happens when agents run long enough to get impatient or socially influenced by other agents, a real concern as multi-agent systems get deployed for longer, less supervised tasks.
It's a small, synthetic test, not evidence that AI has a self-control problem. But stripping personas of hedonic drive and age pushed completion rates close to 1.0, which says more about how sensitive these agents are to prompt framing than about anything resembling willpower.