5 min readAI Agents
How do you score whether an AI agent's actions can be undone?
By Josh Woodruff, Founder & CEO, MassiveScale.AI | Founding Chair, Agentic Trust Framework at the CSAI Foundation

TL;DR: List every action your highest-privilege AI agent can take. Give each one an undo score from zero to two. The agent reversing it scores two, and a person reversing it with a known procedure scores one. Zero means nobody can. Any zero with a big impact needs a human second key. Then prove it with one timed restore drill.
I spent three days at the AI Engineer World's Fair in the summer of 2026 and came home with 584 slides.
One idea ran through the best of them. The top builders have stopped trying to make agents trustworthy. They limit what an agent can touch, and they assume a bad action gets through anyway.
That changes the audit question. It used to be whether you trust the agent. Now it's whether you can undo what the agent did last night.
What is the undo test?
It's two questions you ask before an agent gets any permission.
Can the action be reversed?
How much breaks if it's wrong?
If either answer is bad, a human holds a second key. The agent can own a small test rollout. A person owns production.
The question "do we trust this agent" has no answer you can check. The question "can we reverse this action" does. I lay out the thinking on MassiveScale.AI, in can you undo what your AI agents did last night.
How does the assessment score this?
Question 29 of the free assessment asks whether you can recover from agent incidents. It's part of Incident Response, one of the five ATF elements. Each answer puts a different ceiling on your undo scores.
Answer | What it says | What it means for undo |
|---|---|---|
A | No recovery procedures for agent incidents. | Every write action scores zero until proven otherwise. |
B | General system recovery only. | You can restore a whole system, and lose everyone's work since the backup. |
C | We can restart agents and restore basic operation. | The agent comes back. What it changed stays changed. |
D | We can roll back agent state to a known-good checkpoint. | A person can undo the agent's work to a set point. |
E | Full checkpoint and resume, with a step down to lower autonomy. | The agent's work can be rewound, and it comes back with less freedom. |
How do you score one agent's actions?
Start with the agent that can do the most.
List every action it can take. Use its tool list. Don't work off the job description.
Mark each one read or write. Reads don't need an undo.
Give each write an undo score. Use the table below.
Rate the impact of each write. Small or big is enough.
Find every zero with a big impact. Those need a human approval in front of them.
Undo score | What it means | Example |
|---|---|---|
2 | The agent can reverse it by itself | Moving a ticket to a new queue |
1 | A person can reverse it with a known procedure | Restoring a record from yesterday's backup |
0, small impact | Nobody can reverse it, and little breaks | Sending an internal status note |
0, big impact | Nobody can reverse it, and a lot breaks | Sending money or emailing a customer |
The pass bar is one rule. No zero with a big impact runs without a person saying yes first. I show how to prove that approval step holds in testing whether you can stop an AI agent before it acts.
How do you prove a score of one or two?
Run a restore drill. A score on paper is a claim.
Pick one action that scored a one. Have the agent do it in a test setting. Then hand the procedure to someone who didn't write it and start a clock.
If they can't finish, the score was really a zero. If it takes four hours, write that down. Four hours of wrong data is part of your impact.
Put a rate limit on writes while you're there. Reads can be generous. Writes shouldn't be. A slow agent gives you time to catch a mistake before it spreads.
Why score actions and not the model?
Because the actions are the part you own.
A better model doesn't make a wire transfer reversible. The rules around the model do. They decide what the agent can touch, and they keep the record of what it did.
The same week, Amplify Partners presented its 2026 State of AI Engineering survey of 1,048 respondents. The most common way teams check their AI is still an informal vibe review. Gut feel. An undo score is the opposite of that. It's a number per action, and anyone can check it.
Frequently asked questions
Does every action need to be reversible?
No. Some never will be. A sent email stays sent. Those actions need a person in front of them.
Who should assign the scores?
The agent's owner and the team that runs the system it writes to. The owner knows what the agent does. The system team knows what a restore really takes.
How often should I rescore?
Each time the agent gets a new tool. One new write action can add a zero.
What should be true before an agent goes live?
It should pass both halves: you can shut it off, and you can undo it. Michelle covers that in what evidence proves an agent passes the shut-off and undo test.
Key takeaways
Score every write action from zero to two, by who can reverse it.
Any zero with a big impact gets a human second key.
Prove a score with a timed restore drill, run by someone who didn't write the procedure.
Rate-limit writes so mistakes spread slowly.
Question 29 is one of 30 in the free assessment. It takes about ten minutes and scores you on all five ATF elements.
The best builders I know gave up on trustworthy agents. They kept the receipts and the undo.