Don't trust large context windows

Large language models lose effectiveness well before reaching their advertised context window limits, with a practical "smart zone" cutting off around 100k tokens despite vendors marketing windows of 200k to 2M tokens. Coding agents rapidly consume tokens and enter this degraded "dumb zone," where the model forgets previous instructions, and while auto-compaction helps, the better approach is deliberately moving information into external artifacts between sessions to keep the working context sharp and efficient.