A team of AI agents just solved a messy physics formula that stumped most AI agents working alone.
Researchers tested how AI systems rediscover scientific formulas, using the "hydrotope" as a case study: a geometric formula that describes how nonlinear water waves scatter, stitching together different polynomial expressions depending on the frequency region involved. They ran 18 single-agent attempts, with no hints, a false hint representing a wrong assumption, and a true hint representing real domain knowledge. Only four of those runs recovered the complete formula across every region. Most agents found the right polynomial for a given region but couldn't figure out how to combine the pieces or confirm the result held everywhere. Standard symbolic regression tools, including LLM-assisted versions, and conventional machine-learning regressors failed too.
The fix wasn't a smarter model, it was a different org chart. The researchers built a workflow with one "lead" agent that assigned analytic and numerical sub-tasks to two "student" agents and independently checked their work. That setup rediscovered the full hydrotope formula, then went further: applied to a harder variant of the problem, it derived a new, independently verified formula for a six-point wave amplitude that hadn't been published before.
It's one case study on one wave-physics problem, but the failure pattern matches other agent benchmarks: solo agents stall on problems that require combining partial results, while splitting the work across roles with a supervisor checking the output does better.