Improve the Process That Proposes Improvements

Abstract

An agent can accumulate better task routines while its ability to discover the next improvement stays fixed. We propose a lineage experiment that allows research agents to modify a bounded policy for proposing and testing their own harness edits. At each generation we will transplant the evolved mutation policy into a common fresh executor to measure its ability to discover improvements without inherited task routines. The goal is to determine whether the improvement process itself becomes more sample-efficient under a fixed resource budget.

Keywords