Can I change a specialist’s instructions without a redeploy?
Ask this by asking for something instead. Somewhere in the demonstration a specialist will give advice that is competent and wrong for your company: a discount recommended where you hold list price, a tone you would never use with a client. Say so, and ask them to change it now, while you watch. In about four minutes you learn whether the judgement inside this system is your policy or their product.
Text, or something that has to be built
A specialist is a name, a model behind it, and a set of written instructions about a screen long. The question is when those instructions get read.
In the Cabinet, ArkOne’s reference design for an executive agent, each of the ten specialists reads its instructions at the moment it is consulted, fetched then from wherever your company keeps them. Change the text tonight and tomorrow morning’s meeting answers under it. Nothing is rebuilt and nobody outside your company is asked. The Register, the standing page of every setting in the design, treats those instructions as exactly that: a setting.
The machinery around them is a different matter. The loop, the rules that decide what may leave the room, the ledger: those are code, changed on a release schedule with an engineer behind it. That separation is the point. Machinery gets built once and rarely touched. Judgement moves whenever your policy moves, which in a company of your size is most months.
Four minutes, five things
Suppose you have asked for that one line, in a demonstration invented for the exercise: never propose a discount to close a renewal, offer payment terms or a volume tier instead.
| Minute | What they should be able to do | What it tells you |
|---|---|---|
| The first | Open the specialist’s instructions on screen and let you read them | Instructions you cannot read are instructions you cannot own |
| The second | Add your line and save it, without leaving the product | An edit that opens a support ticket is a release with better manners |
| The third | Re-run the same question and show the advice change | The gap between saving and taking effect is the whole answer |
| The fourth | Show who changed it, when, and the wording before | Without a history, nobody can read a transcript against the rules in force |
| Then | Show the control that puts the old version back | An edit that lands at the next meeting needs a way back in a minute |
Exhibit 1. Illustrative. One line edited during a demonstration, and the five things that happen next in a room built for it.
The third minute is where most demonstrations end. An edit that saves cleanly and takes effect at the next release has changed nothing today, and a room whose policy moves on a fortnightly cadence spends that fortnight advising against decisions you already made.
A careful buyer weighs the fourth and fifth rows above the speed. A sentence takes effect tomorrow morning whether or not it was wise, so the edit is worth what stands between a careless line and a client: a kept previous version, a history with a name on it, and a scored examination the change must survive.
What a prepared vendor shows you
The reply worth having is a screen rather than a description, and it names the risk unprompted. Here are the instructions, the history, the restore control, and the scenarios that run when a line changes. A vendor with all four was burned by an edit once and built the brakes afterwards.
Two thinner replies come up often. The first is that changes go through their team and arrive in the next release, a redeploy described politely; the follow-up is how long, on average, this year. The second is more candid and more expensive: that the instructions have been carefully tuned and changing them risks quality. That names a genuine danger and, in the same breath, tells you the judgement is theirs. Ask whether anybody at your company will be able to change how their own discounts are handled, ever.
For the vendor who passes both, one last question. If I change a line tonight, what runs against it before a client meets the new version, and may I see the last change that failed those tests?
Who owns the sentence
The specialist’s instructions are where your company’s judgement lives: how discounts are handled, what tone a client hears, when to escalate. The rest is plumbing you can rent. Editing a specialist safely, and the habits around the edit, are worked through in the Cabinet paper on changing a seat. What runs against the new wording before anybody meets it is the question about refusing a bad instruction.
Asked plainly
Can I change an AI agent's instructions myself, or does the vendor do it?
Both arrangements are sold. Where the instructions are stored as text your company owns and read afresh each time a specialist is consulted, you edit them and the next run reads the new version. Where they are compiled into the delivered software, changing one sentence means the vendor building and releasing that software again.
What is a redeploy, and why does it matter for AI agents?
A redeploy is the vendor rebuilding and releasing their software so a change takes effect. It matters because business policy moves faster than release schedules. If the sentence describing how your company handles discounts lives inside the delivered code, every policy change becomes a request in somebody else's queue.
What should sit between editing an AI agent's instructions and it giving advice?
Three things, at a minimum. A record of who changed what and when, so a later transcript can be read against the wording in force. A kept previous version, so a poor edit is a one-minute repair. And a standing set of scored test scenarios run against the change before it meets a customer.