Keep decisions and assumptions separate

The central feature of my chief-of-staff agent was its state management: it kept my explicit decisions separate from its own assumptions.

Whenever I answered a question or gave direction, that went into one file. When the chief-of-staff made a call on its own, or noticed a subagent doing so, that went into another.

That distinction mattered when I gave the agents more room to work. Otherwise, a project could drift into a state I did not know about or had not intended.

The point was to make the difference visible: what I had decided, and what the agents were assuming.