The Book of Works
The agentic workflow
From chat to colleague: plans, permissions, and letting it drive while you steer.
Working with an agentic machine is a different job from chatting with one. You stop dictating keystrokes and start running a colleague: brief it, gate the plan, let it drive, and hold the checkpoints. The unit of work is the session, opened with a bounded goal and closed with something verified, and your effort concentrates at the two ends while the machine owns the middle.
The plan gate is where the leverage lives. A session that begins with “Propose changes first, before editing.” costs one reading and catches the wrong turn while it is still a sentence. Permission modes are the same lever on the action side: decide up front what the machine may do freely, what it must ask about, and what it may never touch, and then let it run instead of hovering.
How to practice it
Open with the gate.
Ask for the approach, review it, then release the build. The sentence costs nothing and converts afternoon-sized mistakes into plan edits.
Set permissions, then stop hovering.
Free for the reversible, ask-first for the risky, never for the forbidden. Supervision belongs in the setup, not over the shoulder.
One session, one outcome.
Scope each session to something you can verify at the end. Sprawling sessions produce sprawling diffs and unverifiable claims.
Close deliberately.
End with verification and an explicit handoff: what landed, what's unverified, what the next session should pick up.