Playbook

How to decide what your agent should handle without asking.

How to decide what your agent should handle without asking.

The hardest part of automation isn’t what to automate. It’s drawing the line between what runs on its own and what waits for you.

woman in black long sleeve shirt

Lorie Jhonson

Head of Customer Success

Read time:

5 min read

Title: Portrait of a Woman, probably a Member of the Van Beresteyn Family,Artist: Rembrandt (Rembrandt van Rijn) (Dutch, Leiden 1606–1669 Amsterdam),Date: 1632,Medium: Oil on canvas,Dimensions: 44 × 35 in. (111.8 × 88.9 cm),Classification: Paintings,Credit Line: H. O. Havemeyer Collection, Bequest of Mrs. H. O. Havemeyer, 1929,Object Number: 29.100.4,Curatorial Department: European Paintings. http://www.metmuseum.org/art/collection/search/437388

Every team we onboard asks the same question in the first week. How much should we let the agent do on its own?

It’s the right question, and there’s no single answer. But after watching hundreds of teams set this up, we’ve seen a pattern in how the ones who get it right think about it.

Start with reversibility, not risk

Most teams start by asking “how risky is this action?” That’s a fair instinct, but it leads to overthinking. A better first question is simpler: can this be undone?

If an action can be reversed in one click, let the agent handle it on its own. A miscategorized ticket, a draft saved to the wrong folder, a status updated too early. None of these are worth a human’s time to approve, because fixing them costs nothing.

The actions worth pausing on are the ones that can’t be taken back.

Draw the line at the point of no return

An email leaving your domain. Money moving. A record being deleted. These are the moments where a Checkpoint earns its place, because there’s no undo button in the real world once they happen.

This is the line that matters. Not “important versus unimportant,” but “reversible versus permanent.” Once you frame it that way, most decisions about what to automate answer themselves.

Let the line move as trust grows

The line isn’t fixed. Most teams start cautious, holding more than they need to, and that’s fine. It’s how trust gets built.

After a few weeks of watching the agent work in the Ledger, teams almost always loosen up. They see that the categorization is accurate, that the drafts are good, that the flags are right. So they let more run on its own, and they keep the Checkpoints only where they truly matter.

That’s the goal. Not maximum automation on day one, but the right amount of it once you’ve seen the work for yourself.

The point was never to remove the human

A good Checkpoint setup doesn’t take people out of the loop. It puts them exactly where they add the most value, on the decisions that are permanent, sensitive, or genuinely need judgment.

Everything else just runs. And you can watch all of it, any time you want to look.

Share this article

//

KEEP READING
AVRA logo mark

© 2026

//

KEEP READING

More from Field Notes.

Get template

Create a free website with Framer, the website builder loved by startups, designers and agencies.