This is interesting.

Do you see a noticeable difference in output quality when the agent reads context first vs going straight into generation?

Feels like most tools skip that step.