Infini-AI Lab’s open project Hermit improves external instructions, skills, memory, tools and code coordination without changing the underlying agent model or harness. It uses task feedback and brings together optimization approaches including GEPA, SkillOpt, MWM and Harness², with exports for Codex and Claude Code. Setup requires Python 3.12 or later, Git and the relevant execution environment. A simple repository demonstration showing a success-rate improvement is a constructed example of the mechanism, not a benchmark proving gains across real development work. Adoption should be evaluated on local tasks and execution cost.