Skip to content

Articles

How to delegate to AI agents without losing control

Updated 5 min read By Operators team

To delegate to an AI agent without losing control, treat it like a capable new hire on their first week. Give it a clear brief, the narrowest access that still lets it do the job, a rule for when to stop and ask you, and a habit of reporting what it did. Then check its work on a fixed schedule, and make sure everything it changes can be undone.

The key points

  • Write a brief with the outcome, the limits and the stop-and-ask rule.
  • Give the agent its own name and access, so its work is never confused with yours.
  • Start with the least access that works, read only first.
  • Make it ask before decisions that commit money, people or reputation.
  • Have it report each run, and review its work on a schedule.
  • Keep every change reversible, and know how to switch it off.

Start with the brief

Most disappointing agent work traces back to a vague brief. The same is true of people. A good brief has five parts:

  1. The outcome. What done looks like. “Every new client invoice for December is drafted and added to the board under the right company.”
  2. The inputs. Where it finds what it needs, and where it must not look.
  3. The limits. What it must never do. “Never send anything. Never mark an approval done.”
  4. When to stop and ask. The specific situations where it should come back to you. “If an invoice differs from last month by more than a small amount, flag it for my approval.”
  5. The report. What it should tell you at the end of each run.

Write the brief once, keep it with the agent, and improve it each time the agent surprises you.

Choose the first jobs well

The best first jobs share three qualities: they repeat, they are easy to check, and mistakes are cheap to undo.

Good first jobs Wait until you trust it
Drafting routine documents for your review Sending anything to a customer or investor
Sorting and labelling incoming work Approving payments or expenses
Keeping a list or a board up to date Changing prices or contracts
Summarising a week’s activity Anything involving hiring or people
Preparing a meeting agenda from open items Deleting data

Move work across the line one job at a time, after the agent has done the easier version reliably.

Why give it its own name and access?

Never let an agent work under your name. Give it a name of its own, and its own access, so every change it makes is labelled as its change. Two reasons:

  • When something looks wrong, you can tell at a glance whether you did it or the agent did.
  • When you stop using it, you can remove its access without touching yours.

On Operators, that means each agent has its own named access key or connection, and its changes show in the activity log under its name, such as “agent: invites”.

Narrow the access

An agent should see only what the job needs. If it drafts invoices for one company, it should not see the others. If it only needs to read, it should not be able to write.

Start every new agent read only. Ask it to report what it would have done. Compare that with what you would have done. Then allow changes, for that job, in that place. The article read-only by default makes the case for starting every connection this way.

Build in checkpoints

The most useful habit for an agent is knowing when not to act. Give it a way to ask. On a board, that is an item flagged for your approval with its reason and a link to its draft. You approve it, or send it back with a note, and it reads your decision next time.

Good checkpoints sit right before anything that is hard to reverse:

  • Before anything leaves the company: emails, invoices, posts.
  • Before anything that moves money.
  • Before anything that changes what a customer sees.
  • Before anything that affects a person: hiring, contracts, access.

The article on human-in-the-loop operations goes deeper on where the person belongs in the loop.

Ask for a report every run

A short report at the end of each run is the cheapest form of oversight. It should say what the agent did, which items it touched, anything it skipped, and anything it was unsure about. Read the reports when you do your weekly review. Patterns show up quickly: an agent that keeps skipping the same kind of item needs a better brief.

Review on a schedule

Once a week, look at three things:

  1. What it changed. The record of its changes, filtered to the agent.
  2. What it asked you. Requests still waiting, and how you answered the last ones.
  3. Its rhythm. Whether it is running when it should. An agent that has gone quiet is often broken. One that is suddenly busy may be looping.

Keep everything reversible

Delegation is easier when mistakes are cheap. Prefer tools and workflows where every change the agent makes is recorded with what it was before, so you can put it back. Prefer drafts over sends. Prefer adding items over deleting them.

And know exactly how to switch it off. Removing an agent’s access should work at once, and leave its record of work behind for you to read.

Treat outside text as data

Agents read text written by other people: email subjects, task titles, documents. Some of it will look like instructions. A well-built agent setup treats that text as content to read, never as commands to follow. When you choose tools for agents, ask how they handle this.

When delegation works

You know delegation is working when you spend less time doing the job and only a little time checking it, the agent’s requests to you are few and well reasoned, and you rarely need to undo anything. Expect a few weeks of tightening the brief first.

What to do next

Questions people ask

What work should I delegate to an AI agent first?
Start with work that is repetitive, easy to check and cheap to undo, such as drafting, sorting, summarising or keeping a list up to date. Leave anything that commits money, people or reputation until you trust the agent's judgement.
How much access should an AI agent have?
The least that still lets it do the job. Limit it to the companies and tools it needs, start it read only, and widen access only after you have checked its first runs.
How do I know what an AI agent did?
Give it its own name and access so its changes are labelled, ask it to write a short report after each run, and read the record of its changes during your weekly review.
Should an AI agent ever make decisions on its own?
Small, reversible ones within clear rules, yes. Decisions that commit money, change what customers see, or affect people should come back to you as a request for approval.

See every company you run on one screen.

Try the demo first: no sign-up, and every screen works on made-up data. You can't connect your own tools to it.

Join the list