Loop

The article argues that making coding agents truly project-aware requires two complementary mechanisms: an ontology that explicitly encodes domain concepts, relations and validity conditions so models understand team conventions, and recursive self-improvement (RSI) that turns execution feedback into reusable knowledge, with memory handling storage and reuse. The authors ground this in their internal Loop practice, aiming to cut the time agents spend re-learning a project and to stop them repeating mistakes the team has already resolved.