The autonomy ladder: place every action deliberately

Autonomy is not one switch you set for the whole product. It is a ladder, and every action sits on its own rung, with its own answer to how much it can do without a human.

The five rungs, from suggest to act silently

What separates one rung from the next is what the human still does on it.

• Suggest. The product proposes and the human acts. A research tool that recommends which contracts to read first sits here, beside a support console that proposes the macro for the ticket.
• Draft. The product produces the finished artifact and the human edits and sends it. The reply exists in full; the send button stays human.
• Act with approval. The product is ready to execute and waits for an explicit yes. The human stops doing the work and starts gating it.
• Act with undo. The product executes on its own, and the human can reverse the effect within a window: restore the archived thread, release the hold, roll back the record.
• Act silently. The product executes, and the human moves after the fact, reading receipts and auditing samples.

No rung removes the human entirely. Even the top one keeps a person in charge, just later in the process, the arrangement Supervision: keep a human in charge of the agent argued every agent needs. If nobody ever reads the receipts, the action has fallen off the ladder rather than climbed above it. Graduated autonomy is not our invention either: it carries about half a century of human-factors research, built on scales from the person doing everything at one end to the machine acting and telling nobody at the other.

Two questions place each action: how reversible, how far-reaching

You place each action on a rung, never the product as a whole. Ask two things about it: how reversible its effect is, and its blast radius, meaning how much it touches before a human can react. The harder it is to undo and the further it reaches, the lower its rung. A wrong label costs one click, a wrong reply to a single customer costs an apology, and that same reply sent to a whole list can cost you a press cycle. The calendar hold and the payment never share a rung, however well the model handles both.

• Easy to undo and contained actions, like labeling, sorting, and archiving inside your own product, can sit high from the start, often at act with undo in the first week.
• Hard to undo actions, above all messages that go out to people, start at draft or approval however polished the drafts read, because nothing unsends an email a customer has already opened.
• Impossible to undo or far-reaching actions, such as moving money, deleting data with no backup, or writing to systems your user does not control, start at suggest or approval and climb slowly, if they climb at all.

The inventory you built in The action surface: every tool is delegated authority is your starting list, and every action on it gets its own rung. Watch for one tool that hides several actions inside it: replying to a teammate and replying to a customer list may be the same function with different arguments, yet they belong on different rungs.

Place each action one rung below where you think it belongs

Wherever your reasoning lands an action, drop it one rung lower, because the two ways of being wrong do not cost the same. Place an action too low and you pay in extra clicks and friction; place it too high and you pay with an incident, which burns more trust than a month of smooth approvals earns back. We start every new action below our own confidence, and it has never felt slow, because moving one back up is a routine step we take ourselves rather than a fight we have to win.

To promote an action up a rung, you need two kinds of evidence together. On the eval side, its cases pass at the bar, scored by the Graders: deterministic, judges, and humans you calibrated and held in place by The regression gate: no change ships blind, so a later change cannot quietly drop the pass rate underneath it. On the production side, it has run a set stretch at real volume, with no incident traced back to it. Write each promotion down as a dated decision. Demotion needs no meeting: when an incident happens, move the action down a rung first and investigate second, because the ladder stays honest only if it moves both ways.

Too many approvals, and people stop reading them

The approval rung looks like the safe default, and that is exactly why it fails. Decades of human-factors research describe automation complacency: when people supervise automation that is usually right, they stop inspecting it, and a click that began as a real judgment turns into a reflex they perform while reading something else. A gate users click through without looking offers no real control, and it is more dangerous than an honest act-with-undo rung, because the log now claims a human approved something nobody examined. Keeping the rung real comes down to what the approval shows and how often it fires.

• Show a real preview. The approval has to show the action itself, not a summary of it: the full email under the recipient's actual address, the exact amount and payee, the precise records a cleanup will touch. Anxiety: lower the stakes at risky moments showed how a real preview makes a risky moment manageable, and the approval prompt is that same move.
• Keep approvals rare. If most actions fire an approval, users are trained out of reading them within days. Move cheap, reversible actions up to act with undo and unreliable ones down to draft, so the approvals that remain are rare enough to still get read.

Measure the gate, too. When the approval rate sits near total and the time from prompt to click drops to a second or two, the gate has stopped working, and that is a signal to redesign it, not trust you have earned.

Try it now

The drill takes about fifteen minutes and runs on the action inventory for your own agent feature, real or planned.

Place every action on exactly one rung. Write the rungs from suggest to act silently and place each action from your inventory by reversibility and blast radius. Whenever you are torn between two rungs, take the lower one; that is the starting placement, not the final word.

Write the evidence sentence for everything above act with approval. For each action at act with undo or act silently, write one sentence naming the eval result and the production record that justify the height: "passes its cases at the bar and has run a month at approval, at full volume, without an incident."

Demote what has no sentence. If the sentence will not come, or it begins with "we believe" or "it should", the action moves down a rung today. You are not punishing the action; you are pricing the missing evidence.

Make the top rungs defend themselves. Paste the placed ladder into Claude Code and ask it to argue every action above approval down one rung, naming the worst plausible turn at its current height. Keep every demotion it wins, and turn the arguments you reject into eval cases, because each one names evidence you now owe.

Chapter Summary

• Autonomy is not one setting for the whole product; each action earns its own level of independence, one rung at a time.
• The rungs run from suggest, to draft, to act with approval, to act with undo, to act silently, and what changes between them is how much the human still does.
• No rung removes the human, so an action whose results nobody ever checks has fallen off the ladder rather than topped it.
• Place each action by how easily its effect can be undone and how far it reaches; the harder and further, the lower the rung.
• You place actions, not the product, so the calendar hold and the payment land on different rungs even though one model handles both.
• Start every action one rung below where you think it belongs, because landing too high costs an incident while landing too low only costs some friction.
• Promote an action only once its eval cases pass and it has run a real stretch at full volume with no incident; demote the moment one happens, and investigate after.
• Keep approvals rare and make each one show the real action, because an approval people click through without reading is worse than no gate at all.
• Next up is Blast radius: bound what one turn can touch, which shows how to shrink the reach of a single action.

Sources

• Sheridan, T. B., & Verplank, W. L. (1978). Human and Computer Control of Undersea Teleoperators. MIT Man-Machine Systems Laboratory.
• Sheridan, T. B. (1992). Telerobotics, Automation, and Human Supervisory Control. MIT Press.
• Parasuraman, R., & Riley, V. (1997). Humans and Automation: Use, Misuse, Disuse, Abuse. Human Factors, 39(2).
• Parasuraman, R., & Manzey, D. H. (2010). Complacency and Bias in Human Use of Automation: An Attentional Integration. Human Factors, 52(3).