How a Silent Context Window Overflow Made an AI Agent Forget Its Instructions

A developer discovered that their AI refactoring agent began making errors around turn 37 of a session — not due to model degradation, but because the context window had silently filled up. The agent's framework was evicting the system prompt to free space, causing later turns to run without the original instructions. Debugging revealed the framework's token estimator only counted user and assistant turns, ignoring the system prompt, tool schemas, and few-shot examples, causing it to underreport actual token usage by roughly 12 percentage points. The developer had been testing on MonkeyCode's free server with a 10-million-token monthly allowance, which created a false sense of unlimited headroom and led to neglecting per-session context budgeting. The incident highlights that a large monthly token pool does not expand a model's fixed context window, and developers must track all token consumers — not just conversation turns — to avoid silent instruction loss.