Incorrect atomic and lock-free synchronization changes can cause rare, architecture-dependent crashes or data corruption that evade testing and are exceptionally difficult to reproduce and diagnose. WT-10461 affected 14 customers and took years of engineering effort to resolve. Changes made without sufficient understanding, including reliance on plausible but incorrect AI reasoning or safety claims, increase the risk of serious customer-impacting defects.

Add rules to AGENTS.md prohibiting independent AI-agent correctness analysis and allowing atomic or lock-free changes only as explicitly specified by a developer. Prefer established locks for new synchronization, and clarify that atomics-backed mutexes remain permitted.