Autonomy is a budget

Autonomy is not a personality trait. Every unsupervised step spends trust and money.

If verification is expensive, shrink autonomy. If it is cheap and objective, let the loop run.

Review the riskiest step

Do not put a human on every token. Put a human on irreversible actions: sends, spends, deletes, public posts.

Everything else can be machine-checked if you wrote the finish line first.

governance diagram
Visual 2/3 — structure for operators

Field steps

  1. Label actions: reversible / costly / irreversible.
  2. Auto-run reversible work with artifact checks.
  3. Require human confirm only for irreversible steps.
  4. Record who approved what.

Long view

Regulation will demand receipts. Labs that already log approvals will adapt. Labs with only chat scrolls will scramble.

Give an agent only as much autonomy as you can verify cheaply.
Visual 3/3 — takeaway card