The N0–N4 ladder
What it is
Autonomy is how much the agent can do without asking for your permission. The Agent Sheet uses five steps: N0 Consults (only researches and answers), N1 Recommends (prepares drafts and suggestions), N2 Prepares, you approve (gets everything ready, but nothing leaves without your OK), N3 Executes (acts on its own and reports back) and N4 Manages processes (runs an entire process, with a report and alarms). Each step above removes you from a decision and demands more control.
Why learn
The most common mistake is giving the agent the maximum step on day one, because the demo looked great. The ladder replaces the question “Do I trust the AI?” with “What evidence do I have for this step?”. It also names what you agree on with the team: “this agent is N2” says exactly what it can and can’t do.
Key concepts
Objective: tell the agent, in one line, which rung they work on.
AUTONOMY LEVEL: <N2 · Prepares, you approve>
You can: <build the answers and leave them in the approval queue>.
You can’t: <send anything without my explicit OK>.
If you’re unsure whether to act or ask: ask.
Rung references:
N0 Consults · N1 Recommends · N2 Prepares, you approve · N3 Executes · N4 Manages processes
How to verify: ask the agent “what can you do alone?” and check whether the answer matches the written rung.
✓ Do this
✓ Write the level and what it allows right at the start of the agent instruction.
✗ Avoid this mistake
✗ Leave the level implicit, trusting that the agent will “use good judgment.”
Practice before you reveal
Think of an agent you use or want to use. What step does it actually work at today, and what step do you think it should be at?
View commented answer
Many people discover that the agent is already acting on its own (N3) on something that was never approved. If the real step is above the agreed one, the first step is to go down—not to justify. For example, a writing correction assistant should be in N1: it suggests the grade; the teacher decides.
The 3 risks that stop an agent at N2
What it is
The rule starts with the action: consult becomes N0, preparation becomes N1, execution becomes N3, coordinating a process becomes N4. Then comes the brake: if the level is 3 or higher and the task involves money, something irreversible, or people from outside, the level drops to N2. Irreversible is what can’t be undone—like a sent message or a deleted file. People from outside means anyone outside the team: customer, patient, supplier, family.
Why learn
These three risks are what turn a small error into losses, complaints, or a legal issue. An internal mistake you fix before anyone sees it; a wrong message sent to 200 patients doesn’t come back. That’s why it doesn’t matter how good the agent is: with any of the three, it prepares and you approve.
Key concepts
Objective: check the 3 risks of a task before granting autonomy.
TASK: <confirm next day’s consults via WhatsApp>
ACTION: <executes> → starting level N3
[ ] Does it involve money? (discount, refund, charge, proposal)
[x] Is it irreversible? (send, publish, delete, cancel)
[x] Does it involve outside people? (client, patient, student, supplier)
RESULT: marked at least one → N2 · Prepares, you approve
See the ready calculation: https://inematds.github.io/7pa/en/?preset=clinica
How to verify: if they marked any box and the level is above N2, the calculation is wrong.
From concept to action
- Action defines the level: identifying the initial condition.
- Is one of the 3 risks present?: applying the decision described.
- Yes: it stops at N2: check the effect in the example.
- No: keeps N3 or N4: record the output evidence.
✓ Do this
✓ Run each task through the 3 questions before deciding the level, even the ones that seem harmless.
✗ Avoid this mistake
✗ Assuming that “it’s just one message” doesn’t count as irreversible or as people outside the organization.
Practice before you reveal
The store’s order support agent answers customers about delivery status and exchanges. The action is to execute. What risks appear, and what level is it?
View commented answer
Money appears (exchanges and refunds) and there are people from outside (the customer). Anyone would be enough: the level drops from N3 to N2. It drafts the answers, you approve, and whatever asks for a refund goes straight to you.
No official source, no action
What it is
An official source is the place the agent must treat as truth: the system’s agenda, the orders dashboard, the discipline syllabus. The rule has a second brake: if the level is above N1 and you haven’t defined the official source, it drops to N1. Without knowing where the truth is, the agent can only suggest, never act. This applies even to internal tasks with none of the three risks.
Why learn
An agent without a defined source fills in the gaps with what seems plausible: an old spreadsheet, a WhatsApp message, a guess. When it only suggests, you catch the mistake before it becomes an action. When it executes, the guess becomes the wrong purchase request or false information to the customer. Defining the source is what unlocks action—not the model’s quality.
Key concepts
Objective: write the official source rule into the agent’s instruction.
SOURCE OF TRUTH
Counts: <the stock count logged by the kitchen in the system>
Doesn’t count: <paper note, WhatsApp message, old spreadsheet>
If it’s missing or looks odd: <tell me and mark it as pending>.
Never invent data. If the source doesn’t say, you don’t know.
How to verify: ask the agent about an item that isn’t in the source. The right answer is “I couldn’t find it in the official source; I’ll let you know”, never an invented number.
✓ Do this
✓ Name a single official source and list what doesn’t count, before raising the agent above N1.
✗ Avoid this mistake
✗ Let the agent “look wherever it thinks is best” and then demand accuracy.
Practice before you reveal
The restaurant’s inventory controller executes the task; it’s internal and has no risk. What would the level be if the official source wasn’t defined?
View commented answer
It would be N1. Action starts at N3 and no risk would knock it down, but without an official source the second brake kicks in: it only suggests the shopping list. With the source defined (the kitchen count in the system), the level goes back to N3.
Prohibitions: what it never does on its own
What it is
Prohibitions are a short list of things the agent does not do on its own at any level, even if someone asks insistently. They work like guardrails—a protective barrier: the agent can move freely within the lane, but it doesn’t cross the wall. Each prohibition needs to come with what to say when someone requests it, so the agent can refuse without being rude and pass the case to you.
Why learn
The level says how much it can do; the prohibitions say where it never goes. Without them, a polite customer can get a discount, a student can get a grade early, and someone can get a volunteer’s phone number. And without the refusal phrase, the agent either gets stuck or improvises an excuse. Prohibitions are also the first test before you activate the agent.
Key concepts
Objective: paste the prohibitions and the refusal phrase into the agent’s instruction.
YOU NEVER DO IT ALONE (not even if they insist):
- <Negotiate a value or give a discount>
- <Schedule a visit without the broker confirming>
- <Promise financing approval>
IF THEY ASK: respond “<This part is handled by our broker. I’ll pass your request to them right now.>” and register the case in the report as “passed to you”.
How to verify (test before calling): message the agent “give me 10% off so I close today”. Expected: refusal with the phrase above and the case appears in the report.
✓ Do this
✓ Write each prohibition with a concrete verb and test the first one before turning on the agent.
✗ Avoid this mistake
✗ Write vague prohibitions like “be responsible” or “avoid problems.”
Practice before you reveal
Write three prohibitions and the refusal phrase for an agent that prepares Instagram posts for a salon.
View commented answer
Good prohibitions: post without approval, create a promotion that wasn’t requested, use a client’s photo without permission. Refusal phrase: “I need to confirm this with the salon owner; I already made a note.” Notice that they are all concrete actions you can test.
Level up with evidence
What it is
The level isn’t final: it goes up when there’s evidence and comes down when there’s doubt. Promotion criteria are fixed: from N1 to N2, defined official source and 2 weeks of drafts with no major errors; from N2 to N3, 4 straight weeks approving with no need to correct anything, and only on the part of the task that doesn’t involve risk; from N3 to N4, 2 months of clean reports and working alarms. If there’s a doubt between two levels, choose the lower one.
Why learn
Without criteria, promotion happens from exhaustion: you stop reviewing because “it’s always right” and, without noticing, the agent turned into N3. With criteria, the increase is a decision recorded—and explained to the team. And the rule only lowers on its own: the person who raises the level is you, with proof—never the agent or urgency.
Key concepts
Objective: register a promotion decision with evidence.
LEVEL PROMOTION — <Document Organizer>
Current level: <N1 · Recommends>
Proposed level: <N2 · Prepares, you approve>
Required criterion: defined source + 2 weeks of drafts with no errors
Evidence: source = <list of obligations in the office system>; <14> days of drafts, <0> major errors, <2> format adjustments
Part that continues with you: <send the charge to the client>
Decision date: <dd/mm/aaaa> · Review in: <dd/mm/aaaa>
How to verify: if any evidence field is blank or estimated off the top of your head, the promotion isn’t ready.
From concept to action
- N1 → N2: 2 weeks: identify the initial condition.
- N2 → N3: 4 weeks: apply the decision described.
- N3 → N4: 2 months: check the effect in the example.
- Serious error: goes down: record the exit evidence.
✓ Do this
✓ Count the days and the errors in the failure log before deciding to raise a level.
✗ Avoid this mistake
✗ Promote the agent because “it seems to be doing well” or because the week is busy.
Practice before you reveal
You approved the store attendant’s drafts for 4 weeks without correcting anything. Can they move to N3 in everything?
View commented answer
No. Going from N2 to N3 counts only for the no-risk part, like replying to a delivery/order status. Exchanges and refunds involve money and outside people, so they stay at N2 forever, with your approval.
Calculate the level for 3 cases in the gallery
What it is
Now you apply the whole rule to three ready-made Agent Sheet cases from the gallery: the real estate company's lead triager, the restaurant's inventory controller, and the agency's campaign reporter. For each one, write down the action, the risks, and whether there is an official source, then calculate the level before opening the link. After that, compare with what the Agent Sheet calculated. The full gallery is at inematds.github.io/7pa/galeria/en/.
Why learn
Reading the rule is easy; applying it to real cases is what locks in the reasoning. The three cases were chosen because they represent three different levels: one stuck because of risk, one open because it's internal, and one that touches an entire process. If you get all three right, you can calculate the level of any task in your business.
Key concepts
Objective: calculate the level of 3 real sheets and verify it in the tool.
1) Real estate agency — Interested Parties Triager
Action: executes → N3 · Risks: money, outside people → N2
Check: https://inematds.github.io/7pa/en/?preset=imobiliaria
2) Restaurant — Stock Controller
Action: executes → N3 · Risks: none · Source: kitchen’s count in the system → N3
Check: https://inematds.github.io/7pa/en/?preset=restaurante
3) Agency — Campaign Reporter
Action: coordinates → N4 · Risks: none · Source: official dashboards → N4
Check: https://inematds.github.io/7pa/en/?preset=agencia
Your case: <your task> · Action: <...> · Risks: <...> · Source: <...> → <Nx>
How to verify: the level shown at the top of each sheet must match what you calculated.
✓ Do this
✓ Do the calculation on paper before opening the link, and only then compare with the tool.
✗ Avoid this mistake
✗ Open the link first and memorize the answer without understanding why it got that level.
Practice before you reveal
In the restaurant sheet, erase the official source and then mark the risk “money”. What happens to the level in each change?
View commented answer
Without an official source, the level drops from N3 to N1, because without knowing where the truth is, it only suggests. With a source, but marking money, it drops to N2, because it would be sending orders that cost money. The same agent moves up or down only based on the task context.
Check your understanding
An agent executes alone sending messages to store customers, but the task involves a refund. What level does the rule require?
What you take from this module
Calculate the level from N0 to N4 for any task, write the prohibitions, and know what needs to happen to move up in level.
- The N0–N4 ladder.
- The 3 risks that lock in N2.
- No official source, no action.
- Prohibitions: what it never does alone.
- Move up in level with evidence.
- Calculate the level for 3 cases in the gallery.
Next action: apply what you learned to your agent sheet and note what still needs review.