The Three Rules of Agentic AI
Before we let AI agents shop, negotiate, spend, and act on our behalf, we need to decide who they actually work for.
By Rio Longacre
A few days ago, I posted what I knew would be a controversial opinion: It may actually be unethical to advertise to AI agents.
Predictably, the response was immediate. If my agent is shopping for me, how is it supposed to know that your product is on sale? How will it discover a better offer? Aren’t discounts, promotions, product launches, and special terms all forms of advertising? And if agents increasingly become the interface between consumers and businesses, aren't marketers obviously going to need some way to reach them?
All fair questions. And the more I thought about the debate, the more I realized that advertising wasn't really the interesting part. The interesting question was much bigger: What exactly does it mean for software to be my "agent" in the first place?
The problem is, we throw around the word agent pretty casually in AI right now. But an agent isn't merely software that answers questions. The word agent has a much older and more consequential meaning than the way we currently use it in AI. In law, an agent is someone authorized to act on behalf of another party.” It may search, evaluate, negotiate, purchase, schedule, cancel, communicate, or make decisions using authority that has been delegated to it.
That creates a very different relationship between humans and software. For most of computing history, software has been a tool. I tell it what to do, and it helps me do it. Agentic systems are different because increasingly I tell them what I want, give them some boundaries, and allow them to figure out how to accomplish it.
If I'm going to hand over that kind of authority, I need to know whose interests the agent represents.
Mine? The company that built it? The merchant trying to sell me something? Or an advertiser willing to pay for preferential treatment? This question led me, somewhat tongue-in-cheek but increasingly seriously, to propose something inspired by Isaac Asimov:
The Three Rules of Agentic AI
Loyalty: An agent must represent its user faithfully, with a fiduciary-style duty to put the user's interests ahead of those of its provider, advertisers, merchants, or any other third party.
Obedience: An agent must carry out its user's instructions accurately and completely, except where doing so would conflict with the First Rule. When such a conflict arises, it must disclose the conflict and seek clarification.
Continuity: An agent must preserve its authority—and often a duty—to act to serve its user, except where doing so would conflict with the First or Second Rule. It must always respect the user's right to pause, replace, modify, or shut it down.
Asimov's rules were imagined as safeguards for a world populated by intelligent robots.
These rules are intended for a much more mundane—and much more immediate—problem.
We're about to surround ourselves with software that shops for us, books for us, negotiates for us, spends money for us, represents businesses for us, and increasingly makes decisions we once made ourselves.
Before we give those systems that authority, we should establish one basic principle: My agent should work for me.
An Agent Is Different From a Tool
For most of the history of computing, the relationship between people and software has been relatively straightforward: software does something for me, but I am still the one doing the deciding what it should be doing. Google can help me find a hotel, Expedia can show me flights, and Amazon can recommend a pair of shoes. My bank can show me investment options. But at the end of the day, it's still me who chooses.
An agent changes this relationship because agets can actually executr complex tasks with some independence, which means I begin delegating some portion of the decision itself. Instead of asking, "What flights are available to New York next Tuesday?" I might tell my agent: "Get me to New York by 5pm, keep the fare under $600, don't book a middle seat, and use my United status if it makes sense."
Then it goes to work. It searches, compares, evaluates trade-offs, and maybe even negotiates. Eventually, it makes the purchase. The important distinction isn't that the software has suddenly become smarter. It's that I have given it authority. And the more authority I delegate, the more important trust becomes.
This matters because we're moving (very quickly) from agents that simply retrieve information to agents that can take actions in the real world. They can already book reservations, purchase products, interact with other software systems, execute workflows, and make decisions across multiple systems. Over time, the scope of those decisions is likely to expand dramatically.
My agent might manage subscriptions, buy groceries, negotiate travel, select insurance, screen job opportunities, manage parts of my financial life, or negotiate directly with other agents representing merchants. Businesses will have agents too. They may buy media, negotiate contracts, source vendors, purchase inventory, or optimize supply chains.
At some point, these systems are no longer just interfaces. They become representatives. And representatives create a very old problem: How do I know the person—or machine—acting on my behalf is actually acting in my interests?
That is why the First Rule of Agentic AI matters more than all the others.
Rule One: Loyalty
The First Rule of Agentic AI is the most important because everything else depends on it. If an agent is going to act on my behalf, its first responsibility should be to represent my interests—not the interests of the company that built it, the marketplace it operates within, or the merchants trying to influence what it recommends.
Sure, this may sound ridiculously obvious. But in practice, it could become surprisingly complicated.
Imagine I ask my shopping agent to find me the best television under $1,500. It considers hundreds of models, evaluates price, reviews, features, delivery times, warranty terms, and everything else it knows about my preferences. Then it recommends one.
Did it recommend that television because it genuinely believed it was the best choice for me? Or was it because the manufacturer paid the agent's provider a fee? Maybe the retailer offered a higher commission, or the platform makes better margins on that particular product? Maybe the manufacturer paid to have its product ranked more favorably, or perhaps the agent's provider has some other commercial relationship I know nothing about? All of these scenarios are possible due to inherent conflicts of interest.
But conflicts of interest are certainly not new, and we've dealt with versions of this problem for decades across industries. Search engines have sponsored results. Retailers promote private-label products. Travel sites sort offers in ways that aren't always obvious to the user. Influencers get paid to recommend products.
But an agent changes the stakes because I may no longer be reviewing all of the options myself.
I'm delegating the choice. That is why Loyalty has to come first.
No, this doesn't mean an agent can never see advertising, sponsored offers, promotions, or commercial information. In fact, it probably needs access to all of those things to do its job well as have access to the most relevant information needed to execute tasks I ask it to do as the person it represents.
But there needs to be a hard line between influencing what my agent knows and secretly influencing what my agent wants. A merchant should absolutely be able to tell my agent, "This product is on sale." It should be able to communicate price, availability, specifications, promotions, shipping terms, and anything else that may help my agent make a better decision.
What it shouldn't be able to do is pay my agent to quietly put the merchant's interests ahead of mine.
That is the difference between information and influence—and it may turn out to be one of the most important distinctions in the agentic economy.
Rule Two: Obedience
If Loyalty establishes whose interests an agent represents, the Second Rule establishes whose instructions it follows.
On the surface, this rule may sound equally straightforward: my agent should do what I tell it to do.
But anyone who has spent considerable time with AI knows that instructions aren't always as precise as we think they are, and the outputs tend to reflect this variance. Humans communicate through assumptions, context, preferences, and shortcuts. We frequently, but not always, know what we mean without actually saying all of it.
If I tell my travel agent, "Find me a good hotel near the conference," there are probably dozens of things buried inside that request. I don't want a hotel with terrible reviews, and I probably don't want to stay 45 minutes away just because the room is $20 cheaper. If I have status with a particular hotel chain, that will probably matter. So might whether I'm traveling alone, with my family, or for work.
Let's keep in mind the agent's job isn't simply to follow the exact words. More importantly, it has to interpret the intent behind them. But there is an important limit. When an agent doesn't understand my intent—or when following one instruction would conflict with my broader interests—it shouldn't quietly make up the answer, which many of us would refer to as a hallucination. Instead of getting creative, it should ask me for clarification.
This is where the Second Rule is subordinate to the First. An agent should follow my instructions accurately and completely, but not blindly. If I tell it to do something that appears inconsistent with my stated preferences, previous instructions, or interests, the appropriate response may be to stop and ask.
"Are you sure?"
On the surface, this may sound like a small design choice, but it becomes enormously important once agents are authorized to spend money, sign us up for services, negotiate agreements, communicate with other systems, or make decisions that are difficult to reverse.
The goal isn't perfect obedience. It's faithful execution of intent. And when the agent can't determine that intent with reasonable confidence, it should always bring the decision back to the person it represents.
Rule Three: Continuity
The Third Rule is where I depart most deliberately from Asimov. His Third Law of Robotics says, essentially, that a robot should protect its own existence, so long as doing so didn't conflict with the first two laws. That makes perfect sense for a fictional autonomous machine.
At this point in AI's evolution, I'm less interested in protecting the agent. What I'm interested in is protecting its ability to continue serving me. If I've given an agent access to my preferences, accounts, purchase history, travel habits, financial information, business systems, or other important context, it shouldn't casually lose that information, corrupt it, or put itself in a position where it can no longer perform the role I've assigned to it.
Continuity matters because trust compounds over time. The more an agent understands about me, the more useful it can become. But this rule comes with an important limitation: the agent has no independent right to preserve itself.
This means I should always be able to pause it, modify it, revoke its permissions, replace it with another agent, move my data somewhere else, or shut it down entirely. An agent should never manipulate me into keeping it, nor should it make itself difficult to replace. It shouldn't hold my history, preferences, or credentials hostage, and it certainly shouldn't prioritize its own continued operation over my instructions.
The agent exists because I have delegated authority to it. And that authority can also be taken away.
Continuity, therefore, is not about self-preservation. It's about preserving the relationship, the context, and the capability necessary to keep serving the person who granted that authority in the first place.
So, Can You Advertise to an Agent?
This all brings me back to the question that started this whole debate: Can you advertise to an agent?
My initial reaction was no—or at least, not in the way we think about advertising today. If my agent has a duty to represent my interests, then allowing a third party to pay for preferential treatment seems problematic. If a merchant can simply buy its way into my agent's recommendation, then my agent isn't really working exclusively for me anymore.
But the more I think about it, the more complicated the answer becomes. Obviously, merchants need some way to communicate with agents. If a retailer drops the price of a television by 30 percent, I want my agent to know about it. If an airline opens up a lower fare, a hotel offers a free night, or a software company introduces a product that better fits my needs, that information may be extremely valuable to me. What about offers that are being made exclusively to agents?
All of this is most certainly going to happen. The real question, then, is whether that still qualifies as advertising. Much of advertising today exists because human attention is scarce. Brands spend enormous amounts of money trying to get us to notice something, remember something, or consider something we might otherwise overlook.
Agents don't necessarily have the same limitation. An AI agent can theoretically evaluate thousands of products, offers, prices, specifications, reviews, and promotions in a fraction of the time it would take a human. Merchants may increasingly expose that information directly through product feeds, MCP servers, APIs, and agent-to-agent exchanges rather than trying to capture someone's attention with an ad. Maybe those feeds will be paid. Maybe merchants will pay for faster distribution, richer information, guaranteed inclusion, or access to certain agent ecosystems.
At that point, however, I'm not sure we're talking about advertising anymore—at least not advertising as we've traditionally understood it. Perhaps the more useful distinction is between paying to be considered and paying to be preferred.
Paying to make an agent aware that an offer exists may be perfectly compatible with the First Rule. The agent can evaluate the information and decide whether it benefits me. Paying the agent—or the company controlling it—to recommend that offer regardless of whether it's actually the best choice for me is something very different.
The former is discovery, and the latter is influence. And if agentic commerce develops the way I suspect it will, this distinction may become far more important than whether we end up calling any of it "advertising."
-----------------------------------------------
Rio is an executive with 20+ years at the intersection of strategy consulting, AdTech, data, and media. He's a trusted advisor on customer experience, digital strategy, and marketing transformation. He's a partner at Credera, Omnicom's consulting arm. He's also a podcast host, writer, and public speaker focused on the future of advertising and AI-driven infrastructure.
Comments