Most self-improving agents still refine answers inside a shallow loop;

New research from Univ of Minnesota + Seoul Univ takes another route: keep the improver fixed, feed each new layer the code and traces below it, and deepen only while performance improves.

Instead of letting the agent rewrite the mechanism that improves it, Metan freezes 1 improvement operation and repeatedly asks it to inspect the current solver, its past execution traces, and the improvement code already added, then write a new layer of strategy plus helper functions around the solver.

Each new layer therefore has more evidence about what earlier layers tried and whether it worked, so it can refine or even roll back a bad strategy without risking corruption of the improvement mechanism itself.