New Harvard + MIT + other labs paper shows a cleaner path to self-improving financial AI:

let the agent learn from SEC filing errors, but make every new behavior pass regression tests first.

turns SEC-filing analysis from a static RAG system into one that can continuously learn from failures and safely update how it answers.

It turns repeated mistakes into tested skills, and reject any fix that breaks working behavior.

FINSKILLOPS handles this with a small skill registry. Repeated errors, like using the wrong reporting period or mixing companies, become reusable instructions. A new instruction is deployed only if it fixes the target failure and still passes checks on cases the system already answered correctly.