Why AI Coding Agents Break Their Own Rules Mid-Session and How to Fix It

AI coding agents often stop following user-defined rules partway through long sessions, not because they are incapable, but due to a structural phenomenon called context drift. As a session progresses, tool calls and file outputs pile up in the context window, pushing the original system-prompt rules further into the background where they receive less model attention. Research from Chroma's Context Rot study (July 2025), which tested 18 frontier models including GPT-4.1, Claude 4, and Gemini 2.5, found accuracy consistently declined as context grew, with the sharpest drop between 100K and 500K tokens. A separate 2025 analysis attributed nearly 65% of enterprise agent failures to context drift and memory loss during multi-step reasoning. The recommended fix is to stop relying on static system-prompt rules and instead embed constraints as required, repeatable actions directly before each decision point, keeping the rule recent and contextually close when it matters most.