Let reversible work run and make irreversible work wait

Here is the simplest rule for agent autonomy that works in practice: let reversible work run, and make irreversible work wait for a person. If a mistake can be caught and undone cheaply, the agent may act and be corrected. If it cannot, the agent prepares and someone accountable releases. Most other rules for autonomy turn out to be approximations of this one.

Why reversibility, rather than risk or value?

Firms reach for many measures when deciding how far an agent may act: the amount involved, a risk score, the model's confidence, the customer segment. Each captures something. None captures the thing that matters most when the agent is wrong, which is whether the mistake can be put right.

Value is the most common proxy, and it misleads in both directions. A message to a customer has no monetary value and cannot be unsent. An internal transfer of a large sum between two of the firm's own accounts may be reversed in minutes. A rule based on amount would hold the transfer and send the message.

Model confidence misleads more quietly. A model can be confident and wrong, especially when its sources are wrong. Confidence tells you how the model feels about its answer, not what happens if the answer is mistaken.

Reversibility asks the right question directly: if this turns out to be wrong, what then? The definition of an irreversible action is simple. Applying it takes thought.

What does "reversible" actually mean?

Reversibility is not a yes-or-no property. An action can be technically reversible and practically not. Four questions place an action on the spectrum.

A credit decision is a good test case. The decision can be revisited, so it looks reversible. But the applicant may have acted on it, and other lenders may have seen the search. The article on whether a credit decision is reversible works through why the answer is "less than it looks".

What does "wait" mean in practice?

Waiting is not refusing. An irreversible action that waits is drafted, complete, checked against every bound, and placed in front of a named person who can release it. The agent has done the work. The person supplies the authority.

That distinction matters for adoption. Teams resist rules that stop agents doing useful work. "Make it wait" lets the agent do almost all the work, and keeps only the final, irreversible step for a person. In most processes, that step is a small fraction of the effort.

Waiting also needs to be bounded in time. A held action that sits for days becomes its own harm: a payment that misses its date, a claim settlement that never arrives. Held actions need owners, cover and escalation, just as any queue does.

How do you make more work reversible?

The rule has a useful side effect: it rewards redesigning processes so that more of them are reversible. Many irreversible actions can be split into a reversible part and a much smaller irreversible one.

1. Stage, then release. Prepare payments, messages or changes in a state that has no external effect until released.
2. Add a cooling window. Schedule an irreversible action for later, with a window in which it can be cancelled. Soft deletion before permanent deletion is the classic case.
3. Preserve before overwriting. Keep the old values so a change can be undone.
4. Prefer setting a state over changing it. "Set the limit to 500" is easier to undo than "increase the limit by 200".

Each of these moves work from the waiting side to the running side, without giving the agent any more authority over the step that cannot be undone.

Where does the rule need help?

Reversibility is the main axis, not the only one. Three cases need an extra bound.

Volume. A thousand reversible actions, each wrong in the same way, can do harm faster than a person can undo them. Rate limits and batch checks cover this.

People in vulnerable circumstances. Even reversible actions, such as a reminder or a fee later refunded, can harm someone who is already struggling. Vulnerability signals should lower depth whatever the reversibility.

Weak evidence. A reversible action resting on a forged instruction still rewards the forger. Provenance bounds apply alongside reversibility.

The minimum chain handles this naturally: reversibility sets one bound, and the lowest bound wins.

How do you apply it across an operation?

Start by listing actions and classifying each one. The guide on how to classify actions by how reversible they are gives a method. The list of which AI agent actions are irreversible is a useful starting point, and the irreversible action checklist gives the checks each waiting action should pass before release.

Then set depth by class. Reversible actions can run at depth 3 once certified. Irreversible ones sit at depth 2, or at depth 4 only where a firm has chosen it for a specific capability and earned it on its own cases.

Where Pryme fits

In Pryme, depth follows this rule: depth 3 for routine, reversible work within bounds; depth 4 for moving money and actions that are hard to undo, earned per agent. Actions that must wait return the verdict "held" and go to a named person. See why autonomy in regulated industries is a question about reversal.