SHSam Henry
← The AI Builder Principles

Principle 12

Autonomy is earned

Run the same steps every time, and watch less only after the agent has proven it.

GitHub

Autonomy is a reduction in how many checkpoints along a piece of work the builder personally attends. The path itself does not shorten.

Every unit of work passes five stations in one order. The decision comes first, then the specification, the assignment, the gate, and the verification. Two clauses hold at once. The sequence is fixed. The supervision over it is not.

More autonomy does not shorten the loop. A unit of work run at high autonomy passes the same five stations as a unit run at low autonomy, and passes them in the same order. What changes is how many of the five the builder attends in person, and nothing else changes with it.

Supervision decreases by increments. An increment is bought by evidence from the agent's record of completed work. The builder's confidence buys nothing. The size of an increment is bounded by what the builder is still able to check once the work is done, and that bound is the ceiling on speed.

The sequence is violated by a dropped station, whatever supervision the agent has earned, and by two stations merged into one to save a step. The supervision is violated by an increment granted on the builder's confidence with no record of what the agent produced read, and by an increment larger than the builder is able to check.


Inspection

Walk me through every step a change passes through here, and tell me which ones I have stopped reviewing and what made each of those safe.

Held this, or broke it? Say which principle and what happened — tell me at /contact, subject Principles. What moves a band is somebody else's experience, and I'll publish what moved and who moved it.

SAM HENRY

Founder of SalesSmyth. I build businesses and software with AI agents and write about what holds up. The full version is on About.

New posts reach the newsletter first.

WANT TO COLLAB ON SOMETHING?

I help founders with market positioning, momentum playbooks and building prototypes with AI. A free discovery call is the best way to start.

Grab some time →