New Stanford+Oxford paper MedRSI shows that medical agents can improve themselves from their own mistakes, but only if new capabilities are tested on fresh patients before becoming permanent.

shows self-improving agents need 2 things: focus on harmful failures and refuse to keep new capabilities until they work on later cases.

More important, the paper shows why self-improvement needs guardrails.

If every promising tool is added immediately, accuracy eventually falls: 76.9% by round 30 with 57 tools.

With slower registration, the agent kept just 18 tools and held 94.4%.

It also spends more effort on mistakes that could cause greater clinical harm, rather than just the most common errors.

let agents invent aggressively, but make permanent self-changes earn their place through repeated independent evaluation.