5 min readAI Agents · Cybersecurity
What evidence proves an AI agent is ready to move up a level?
By Josh Woodruff, Founder & CEO, MassiveScale.AI | Founding Chair, Agentic Trust Framework at the CSAI Foundation

TL;DR: An AI agent moves up a level when you can show written evidence for the move. Moving from Intern to Junior takes a clean review record and timed undo tests. The higher moves take 30 days of logs, correct escalations, a named owner, and a review every 90 days. A good demo proves none of it.
Security leaders get squeezed from two sides. Say yes to an agent with no controls, and you own what goes wrong. Say no, and the business builds it anyway.
There's a third answer. Yes, at this level, and here's what earns the next one.
It only works if the next level has a price written on it.
What are the four levels?
The Agentic Trust Framework (ATF) uses four. Think of how you'd bring on a new hire.
Level | What the agent may do |
|---|---|
Intern | It watches and recommends. A person reviews every output. |
Junior | It takes low-risk actions that can be undone, inside its scope. |
Senior | It takes moderate-risk actions, with logging and rules for when to escalate. |
Principal | It works with an earned track record, under regular review. |
Every agent starts as an Intern. A strong demo doesn't change that. The thinking behind the ladder is on MassiveScale.AI, in how to tell what level your AI agents are running at.
What evidence does each move need?
Each move up has its own packet. If a piece is missing, the agent stays put.
Move | Evidence to collect | Pass bar |
|---|---|---|
Intern to Junior | A record of human-reviewed outputs, and a timed undo test for each action it wants to take | Every action can be undone in under five minutes with no data lost |
Junior to Senior | 30 days of action logs with a written error rate, plus examples of escalation | At least two cases where the agent escalated correctly at the edge of its scope |
Senior to Principal | A named owner and logs an auditor could read, plus a review schedule | A formal review at least every 90 days, already on the calendar |
Back down a level | Any trigger from the list further down | The agent earns the move again with fresh evidence |
Notice what isn't in the table. There's no line for how smart the model is. There's no line for how much the team likes it. In the book I wrote with Michelle Savage, a manager named Kevin says it in one line: "Every agent earns its autonomy."
How do you check each piece of evidence?
Each piece comes from a test you can run.
Limits hold. Before any move, show that the agent's current limits work. I cover that in testing whether a new AI agent stays inside intern-level limits.
Actions can be undone. Time the undo for each new action. The method is in scoring whether an AI agent's actions can be undone.
Logs cover 30 days. Open them. Check that every action is there, with what went in and what came out.
The error rate is a number. "It's been fine" isn't a rate. Count the wrong outputs and divide.
Escalations happened. Find two real moments where the agent stopped and asked. If it never has, it hasn't been tested at its edges.
How does the assessment connect to the levels?
The free assessment scores you on the five ATF elements. A move up needs strength in all five, because each level leans on a different one.
Level | The element it leans on most | Why |
|---|---|---|
Intern | Identity Management | You have to know which agent you're reviewing. |
Junior | Incident Response | Its actions have to be stoppable and undoable. |
Senior | Behavioral Monitoring | You're trusting logs and escalation in place of a person. |
Principal | Data Governance and Segmentation | It works with the least oversight, so its data and its reach have to be right. |
If one element scores low, that's the level where your agents should stop climbing for now.
What sends an agent back down?
Trust can be taken back. Write the triggers down before you need them.
A new model or a new tool. The record was earned by a different setup.
An action outside its scope, even one that did no harm.
A missed escalation.
An owner who left and wasn't replaced.
A move down is a reset, and nobody should read it as a punishment. The agent goes back a level and earns the move again with fresh evidence.
How do you use the ladder when the business pushes?
Use it as the deal. When someone wants an agent shipped faster, you aren't saying no. You're saying Intern this week. Junior in 30 days if the logs look right. Senior in 90 if the record is clean. Principal after that, with a review on the calendar.
Most leaders will take that. What they won't take is a flat refusal, and a flat refusal is how security loses its seat.
The first step is getting the agent in the door, which Michelle covers in how to score an AI agent approval request.
Frequently asked questions
Can an agent skip a level?
No. Each level produces the evidence the next one needs. An agent with no Junior record has nothing to show for Senior.
How long does each level take?
At least 30 days from Junior to Senior, because the logs have to cover that long. The other moves take as long as the evidence takes.
Who approves a move?
The agent's owner asks. Security checks the packet. The business owner accepts the risk of the new level.
What does reversible mean here?
Undone in under five minutes with no data lost. If it takes longer, treat the action as one that can't be undone.
Key takeaways
Every agent starts as an Intern, whatever the demo looked like.
Each move up needs its own written evidence and a pass bar.
Junior to Senior needs 30 days of logs and two correct escalations.
Write down what sends an agent back a level.
The ladder turns "no" into "yes, at this level, for now."
The free assessment scores all five ATF elements in about ten minutes. Your lowest element shows where the climbing should pause.
Agents are already inside your company. The ladder decides how far each one has earned the right to go.