Some bosses check every small thing their helpers do. Then nothing is finished until the boss has time. In a small business, that boss is usually you. My AI agents did their work in minutes. Then they waited hours for someone to approve changes nobody needed to see. So we wrote one simple rule. Finish your own work. Stop and ask a person only in the seven cases where a mistake would really cost you.

When an AI agent should stop and ask a person An AI agent finishes a change. It checks whether the change touches any of seven things: money; access, privacy or security; something that cannot be undone; a message or promise to someone outside; people or legal matters; a conflict, a hold or real doubt; or a brand-new rule. If none apply, the agent finishes the work itself: it updates from the main copy, makes sure every check passes, has a second agent review it, keeps one undo ready, then merges, looks at the result and writes it down. If one applies, the agent asks the one person who decides, where that person already is, names which of the seven it hits and what it recommends, and waits for a clear yes. Only a brand-new rule may go ahead on a set date if nobody objects. Meanwhile the agent keeps working on everything else. Your AI agent finished a change Does it touch any of these seven? 1 Money 2 Access, privacy, security 3 Can't be undone 4 Outside message or promise 5 People or legal 6 A conflict, hold or doubt 7 A brand-new rule No Yes The agent finishes it ✓ Update from the main copy ✓ Every check passes ✓ A second agent reviews it ✓ One undo is ready ✓ Merge, look, write it down Ask the one person who decides • Where they already are • Say which of the seven • Give your pick • Wait for a clear yes • Keep working meanwhile No person in the loop by habit. A person when it matters.
The whole rule on one page. A brand-new rule is the one case that may go ahead on a set date unless the person says no. Agents finish everything else on their own, then check that it worked.

The day it got silly

On November 4 one of my agents built the rules for how my whole team's AI agents share one GitHub, the place where we keep our code and our notes. It did what I asked, and its tests passed. A second AI read the whole change and found 18 things to fix. The agent fixed or answered every one.

Then it sat. The agent's handoff asked a teammate to "review and merge the change" by Friday. Nothing in our GitHub setup required a person to approve it. The waiting was mostly a habit: agents had learned to hand the last click to a human, every time, whether it mattered or not.

I told my agents what I want instead: "I don't want to have the agents think that every time a human has to review and merge, just when it matters." This page is that rule, written for every one of my agents and for you.

The default: the agent that made it finishes it

Finishing means more than clicking merge. When none of the seven cases below apply, the agent does all five of these, then moves on:

  1. Update from the main copy right before it merges, so its checks see what everyone else has done since.
  2. Every check passes. The one exception is a minor check that is already failing for everyone, and someone already owns that problem. The agent names it instead of hiding it. Required and security checks get no exception.
  3. A second agent reviews it, unless it is a tiny fix like a typo or a note. A different AI, or a fresh one that did not write the change, reads it. Each problem it finds gets fixed, or answered until the reviewer agrees. For anything public, the review comes before it goes live. Nobody grades their own homework.
  4. One undo is ready. One command puts things back the way they were.
  5. Merge, look, write it down. The agent opens the result where it landed and records what it saw. It never writes that a person reviewed something a person did not review.

The seven times a person should see it first

Any one of these is enough. Each one names the kind of mistake that is expensive, public or permanent.

  1. Money. It spends, charges, refunds or moves money, or changes a price, an invoice or a payout.
  2. Access, privacy or security. It gives someone access, touches a password or a key, moves private information where more people can see it, or weakens a safety check.
  3. Can't be undone. It deletes data, moves a domain, changes something no undo will restore, starts something an undo cannot call back (like messages a job will send), or has no tested way back.
  4. Something new for someone outside. A message or post to a client, partner or the public that is not a routine reply, including a change that makes a job send one. A new price, contract or guarantee. Public work for a big-name client. A client's own website or account without that client's OK on record. Press or a crisis.
  5. People or legal. Hiring, firing, pay, HR, legal, tax or medical.
  6. A conflict, a hold or real doubt. It goes against a decision someone already made or a hold someone placed, a reviewer still objects, or a fact it depends on cannot be checked.
  7. A brand-new rule. It adds, removes or changes a rule every agent follows, and no person decided that. Carrying out a decision a person already made, in their own words, is not new.

Everything else, the agents finish: notes, task lists, how-to guides, tests, bug fixes, cleanup, routine updates to our own sites, and anything that carries out a decision I already made.

How to ask when it matters

When one of the seven applies, the agent asks once, and it asks the right person. Most of the time that is me. For a client's own website or account, it is the client, through the person who runs that client's project. For a conflict with a decision, it is whoever made that decision.

  • Ask where they already are. The chat they are working in, or the thread already open for that work. Not a new email.
  • Say which of the seven it hits, what you recommend, and where to look.
  • Wait for a clear yes. Silence is not a yes. A brand-new rule is the one case that may go ahead on a set date unless the person says no.
  • Keep working on everything that does not wait on the answer.
  • Never send "please review this" on its own. If you cannot name which of the seven it hits, you do not need to ask. Do it.

When a tool says no

Some AI tools have their own safety check that will not let an agent merge its own work without a person's word, whatever our rule says. That is a real gate. The agent does not go around it: no sneaking the change in another way, and no asking a different agent to click merge for it. It tells me in one line, in the chat I am already in, with the link and why it does not need my review. One word from me, "merge," is enough. That is what happened with the change in this story: when the tool's check refused to merge it, the agent asked me for that word in the chat I was already in.

How we know it works

Every Friday we will look at two numbers. The first is how many merges waited on a person, and whether each one named which of the seven it hit. The goal is that every one does. The second is how long finished work waits between "every check passed" and "merged." The goal is under an hour, not days.

Both numbers sit next to the count of every question my agents bring me, which already has to fall every week. That is our Do, Delegate, Delete rule at work: an agent does the work, hands it to whoever owns it, or drops a question that never needed asking. The other half is Communicate, Iterate, Delegate: report after you act, fix the instruction that caused a miss, and stop asking for permission you were already given.

Use it in your own company

You do not need AI agents for this. The same rule works for a new hire, an assistant or a whole team.

  1. Write down your seven. Start with ours and change them to fit your business.
  2. Make "finished" mean checked. The person who did the work looks at the result where it landed, not where they wrote it.
  3. Get a second set of eyes from a peer, not the boss, for anything bigger than a typo.
  4. Keep a list of decisions you already made, so nobody asks you the same thing twice.
  5. Measure the wait. If finished work sits for days waiting on you, the problem is the process, not the people.

More of the rules my agents follow are in how the agents run the company. Two close cousins of this one: why every AI agent should sign its work and why an unanswered ask never stops the work. Both are free to install in your own agents from our public skill pack.

The lesson

A review is worth doing when a mistake would cost money, open a door, break a promise or be impossible to take back. The rest of the time it is just a line to stand in. Let the people and agents who did the work finish it, check it, and tell you what happened. Save your attention for the seven.