A rules engine decides, then checks again.
Agents never hold your credentials. They propose; a rules engine with no AI in it decides, and it runs every rule again when you approve.
Every proposal passes the same checks.
Three proposals, checked one rule at a time, in the order the code runs them.
The checks run in a fixed order.
On October 26, the New Customers agent proposed two raises for the Holiday Blend and gift sets. The engine ran the same checks on both, in order, and the first check that fails ends the run.
| Check, in order | 35%: total $212 to $286 | 8%: total $212 to $230 |
|---|---|---|
| 1. The stop button | 35%: total $212 to $286 Off: passed | 8%: total $212 to $230 Off: passed |
| 2. Budget below zero? | 35%: total $212 to $286 No: passed | 8%: total $212 to $230 No: passed |
| 3. Change to the agent’s total, 20% | 35%: total $212 to $286 35% is over: stopped here | 8%: total $212 to $230 8% is under: passed |
| 4. Daily and monthly caps | 35%: total $212 to $286 Not reached | 8%: total $212 to $230 Passed |
| 5. Largest move without approval, 50% | 35%: total $212 to $286 Not reached | 8%: total $212 to $230 Nothing moves between campaigns: passed |
| 6. Permission level | 35%: total $212 to $286 Not reached | 8%: total $212 to $230 Acts and notifies you: it ran, and you were told |
The stopped proposal still lists every check, with spend limits as the one that failed. No permission level can loosen a limit, because limits run first.
The order the checks run in
- Step 1
Agents can only propose
An agent’s tools read data and propose actions. None of them can change an account. Every proposal says what it would change, from what to what, and why.
- Step 2
The rules run in a fixed order
First, is the action allowed at all. Then the stop button and budgets: none below zero, no change past the size limit, no raise past a cap, large moves sent to you. Ads and messages get contact, audience and content checks. Who approves comes last.
- Step 3
One of three answers, right away
Stopped, with a reason the agent can act on. Needs you, so it waits in your inbox. Or allowed under your permission settings. The same proposal and settings always get the same answer.
- Step 4
Checked again when you approve
Approval can take hours or days, and spend moves in the meantime. So approving runs the policy engine and every guardrail again, against that moment’s spend and settings, before the change is made.
- Step 5
Only the executor holds the keys
One part of the system makes real changes: the executor. It alone holds your account access, and uses it only for a change that passed the rules. Agents never see it.
A 35% raise, stopped and correctedOne example, in order
Oct 26 · New Customers agent
Proposes raising its daily budget 35%, from $212 to $286, all of it for the Holiday Blend and gift sets ($122 to $196): “+$74 a day · +$2,220 a month”. The reason: “The Holiday Blend and gift sets have the lowest cost per first order of any product line.”
Oct 26 · Policy engine
Permission passes: budget changes are allowed. Spend limits stop it: “A 35% change is over your 20% limit per change.”
Oct 26 · New Customers agent
Reads the reason and proposes $122 to $140 instead: an 8% change to its total, inside the limit.
Oct 26 · Policy engine
Every check passes. Budget changes act and notify, so the raise to $140 runs on its own and you’re told.
Logged · Notebook
“Raised the Holiday Blend and gift sets daily budget 15%, $122 to $140, after a 35% proposal was stopped by the guardrail.” Status: ran automatically.
The order the policy engine checks in
Any step can stop a proposal. The level that decides who approves is read last. No level can loosen a limit.
1. The stop buttonCompany setting
When it’s on, every proposal is stopped.
2. Budget below zero?Fixed
A proposal that would set a budget below zero is stopped.
3. Change to the agent’s total, 20%Company setting
By default, an agent’s total daily budget can move at most 20% in one change, up or down. A launch starts from nothing, so the daily cap holds it instead.
4. Daily and monthly capsCompany setting; agent caps only tighter
A raise that would take the daily total, across every agent, over the cap is stopped, and so is any raise once the monthly cap is reached.
5. Largest move without approval, 50%Company setting: 25% or 50%
At the same total, the 20% limit doesn’t apply. Instead, if any campaign moves more than 50% of its own budget, the move waits for your approval.
6. Permission levelCompany setting, per agent
Not allowed stops it. You approve sends it to you. Acts and notifies you and Acts silently let it go ahead.
On every ad and message: contact, audience and contentCompany setting
Consent, suppression lists, contact limits, the daily limit and gap, quiet hours and send days, banned phrases, discounts, figures, required content and brand review’s verdict run on every ad and message at every permission level, and again before anything runs.
When a limit is hit
What happens when an agent argues with a rule?
Nothing. The policy engine is ordinary code with no model in it. The same proposal and the same settings get the same answer every time.
What happens when an approval waits for days?
The rules run again when you approve, against that day’s spend and settings, before the change is made. A proposal that was fine on Monday can be stopped on Thursday.
What happens when two approved changes together would cross a cap?
Each is checked against spend as it stands when it runs, so the second is stopped if the first used up the room.
What happens after a rule stops a proposal?
The stop is logged with its reason, and the agent reads the reason too, so its next proposal can fit inside the rule.
Is there any AI in the policy engine?
No. It’s plain code. Models decide what to propose; code decides whether it runs.
Do the agents ever see my credentials?
No. Agents have no tool that changes an account. Only the executor makes changes, after the rules pass, and only the executor uses your account credentials.
Does the policy engine change my accounts?
No, it only decides. A separate service, the executor, makes the change after the rules run again, and logs the result.
Give agents a real budget. Keep the rules.
Your work email is all we need to start.