Stability is not the same as efficiency, and a new paper argues we have been conflating the two when evaluating AI agents.
Researchers modeled an artificial agent tasked with regulating its own uncertainty, then drove it through a demanding target and reversed course without resetting it. The agent could return to similar internal states either way, but the amount of corrective control it needed to get there was not the same in both directions. The result was a hysteresis loop in regulatory effort — the same destination required different amounts of work depending on the path taken to reach it. Timing mattered too: agents that stabilized before a disturbance hit needed less adaptive effort than those that could only recover after the fact.
This reframes how we should think about agent robustness. A system that looks well-behaved in benchmarks might be burning through hidden regulatory overhead that will only become visible under sustained or compounding pressure — the kind of conditions that benchmarks rarely replicate. For AI deployed in noisy, real-world environments, that gap between apparent stability and actual control burden could matter a great deal.
The broader implication is uncomfortable for anyone shipping agents at scale: passing an output-level test is not evidence that the underlying system is operating efficiently, and two agents with identical performance metrics may have very different margins left before they fail.