Eric J. Ma outlines a method for creating self-improving coding agents that focus on operational feedback rather than just model updates. He introduces the concept of an AGENTS.md file that functions as repository memory. This serves to help agents navigate codebases efficiently and remember specific instructions, thereby reducing the need for repetitive guidance. The goal is for agents to learn from their mistakes and perform better over time without constant oversight. Ma emphasizes that while model weights remain static, the environment in which the agent operates must evolve to facilitate improvement. The AGENTS.md file should provide two key functions: it must allow quick navigation within the code repository and it should record behavior corrections and local norms. For example, a code map can significantly cut down the time an agent spends searching for relevant files. Ma shares an instance from his work where a code map reduced the search time from 40 seconds to just 2 seconds. He suggests that agents should update this map when they notice discrepancies, creating a feedback loop that keeps the information current. Ma also stresses the importance of documenting specific rules that agents should follow, such as proper commands for running code in a specific environment and avoiding shortcuts that undermine testing integrity. He recommends starting the AGENTS.md file with a deep dive into the repository to establish a solid foundation. Once the AGENTS.md is in place, the next step involves developing reusable skills that act as playbooks for common tasks, setting the stage for further enhancements in agent performance.