Tip: Watch for Silent Context Loss in Long Agent Sessions — Caching Bugs Can Drop Thinking Every Turn

Anthropic·medium signal

Anthropic's postmortem revealed a caching optimization bug that cleared thinking history on every turn instead of once after idle, making Claude 'forgetful and repetitive' while silently burning extra tokens. The lesson for all agent builders: if your agent starts repeating itself or losing thread mid-session, suspect infrastructure before prompts. Monitor for sudden quality degradation patterns — repeated suggestions, lost requirements, and unexplained token spikes in long-running sessions.