Blog · 2026-11-02
Earned autonomy, and why your agent should ask less every week
People say control is what they want most from an agent. Then they approve almost every prompt without reading it. In a survey of 16,000 people by Global Payments, 59% of those open to agents said they'd want to approve every purchase one makes. Anthropic found Claude Code users approve 93% of permission prompts.
Both are true. People want control, and they don't want to be asked.
Trust is something you earn
Think about a new assistant. On day one you check every email they send. After a few weeks of good work you tell them to just send the routine ones and copy you in. Then something odd comes up, a new client or a bigger number than usual, and a good assistant checks with you first without being told.
That's earned autonomy. The agent starts by asking, earns the right to act through approvals, and drops back to asking when something is new, big or strange.
Six levels
We use six levels, like the levels for self-driving cars.
| Level | The agent says |
|---|---|
| L0 Suggest | "Here's what I'd do." |
| L1 Ask | "Can I do this?" |
| L2 Heads up | "Doing this in 10 minutes unless you stop me." |
| L3 Act and tell | "Done. Undo?" |
| L4 Act and summarise | "This week I did 14 of these." |
| L5 Autonomous | Nothing, but it's audited and you can always stop it |
Not every action uses every level. An email is either approved or sent with an undo, so a 10 minute delay adds nothing. A payment suits a heads up, because a window to cancel is natural.
Three rules make it work
It's earned per user and per action. An agent can file Sam's receipts on its own and still ask before paying an invoice for Sam on the same day. Trust in one place doesn't leak into another.
Promotion is offered, never silent. After a run of clean approvals the agent asks for more room, in its own words. "You've approved the last 12 of these as they were. Want me to just send them and tell you after?" The human's yes is the promotion.
Anything unusual drops back to ask. That includes a new recipient, an amount well above normal, a new model or prompt, or instructions that came from an email the agent read rather than from you. That one action asks, and the earned level stays.
It has to sit inside your plans and tests
Here's what makes it hard in a real product. Sam might have earned "act and tell" for emails. But Sam is on your free plan, which only allows asking. And Sam is in an experiment arm that holds autonomy back so you can measure the effect.
If your rollout lives in one tool, your experiment in another, your plans in Stripe and your approvals in a function, you get three answers and a bug. The decision has to look at all of them in a fixed order, and say which one decided.
That's what we're building with Throtl. One call answers allow, ask or deny for this user and this action, names the gate that decided, and learns from every answer.
If your agent takes actions for users, I'd love to hear how you handle this today. Join the early access list.