What is a mandate ladder?
The ordered permissions an agent holds over a procedure: advise, propose, apply within a cap, apply.
Definition
mandate ladder (n.) – the ordered set of permissions an AI agent holds over a given procedure, from reading to writing.
A mandate ladder is the rule for how much an AI agent may do on its own in one procedure, and how it earns more. The ladder has four rungs. On advise, the agent reads the system and reports what it finds, including what it would change; it writes nothing. On propose, the agent produces the exact change as a reviewable artifact, a draft, a diff, a queued command, and a person applies it. On apply within a cap, it acts alone up to a stated limit, and anything past that limit drops back to a proposal. On apply, the agent writes to the live system on its own, inside limits that code enforces rather than limits a document requests.
A mandate is held per procedure, not per agent. The same agent may sit on apply for generating a weekly report and on advise for anything that moves money. The rung belongs to the task, not to the worker, so an agent's standing is a list rather than a rank.
Write access is earned by a record of correct proposals and is revoked the moment that record breaks. Demotion is ordinary maintenance rather than a verdict: an agent that starts producing bad proposals returns to advise until the underlying procedure is repaired.
Skipping the propose rung is the common failure. An agent granted apply before anyone has read a season of its proposals is being trusted on an assumption, and the first evidence that the assumption was wrong arrives as a change that has already taken effect. The second failure is quieter: with no proposal history, there is nothing to compare, so no one can say whether the agent's judgment is improving or drifting.
The third failure is a ladder written only in prose. Limits that live in instructions are requests. Limits that live in code are limits. The cap on the third rung is the procedure's blast radius written as a number.
Why the rungs are ordered
The order is not ceremonial. Each rung produces the evidence the next one requires. Advise shows whether the agent understands the system. Propose shows whether its intended changes are correct. Apply spends the trust the first two rungs established. Run in order, the ladder converts an opinion about an agent into a record. Run out of order, it converts an assumption into a live change, and the review that would have caught the error happens after the system has already moved.
“The moat is not the stack. It is knowing which tasks should be fully automated, which need a human checkpoint, and which should never be touched by a model.”
Mandate ladder meaning, with an example
The name is two plain words. A mandate is the authority to act on someone's behalf. A ladder is the order in which that authority is granted: one rung at a time, each earned on the rung below. Asked plainly, the mandate ladder answers one question per procedure: may the agent change this on its own, and if not, how far may it go? The answer is a rung written next to the procedure, not a general opinion of the agent.
A hypothetical: say an agent handles customer emails about late orders. On advise, it reads the inbox and reports which emails it would answer and how; nothing is sent. On propose, it writes each reply as a draft and a person sends it, or edits it first, and the edits become the record. On apply within a cap, it sends replies alone for the one case its playbook covers, a late order with a tracking number, up to a stated number per day, and anything else returns to a draft. On apply, it sends without review, inside limits enforced in code.
Moving up is a decision made on the record, not on a good week: the propose rung shows how often the person changed the drafts, and why. Moving down is the same record read the other way. The failure mode sits at the top: a reply that apologizes for the wrong order looks correct to the agent and to any check that only confirms a message went out. That is why anything sent in the company's name keeps a judgment gate until the verify step can tell a right reply from a fluent wrong one.