AI Agents Got the Keys, and Nobody Changed the Locks

The Spending Problem Nobody Wants to Solve

Here's the funniest headline I've read all week, courtesy of Futurism: bosses are spending more on AI agents than it would cost to just pay human workers. Let that sit for a second. We're in a moment where the pitch for autonomous AI is "efficiency" and "cost savings," and yet the actual math, right now, today, says you'd save money hiring a person. It's like buying a Thermomix to make toast.

But I don't think the economics are really the point, and that's what makes this interesting rather than just absurd. The companies pouring money into agents aren't optimizing for this quarter's payroll. They're buying optionality. They're betting that once these systems work (and yes, I know how that sounds coming from someone who's watched "once it works" promises for two decades in tech), the marginal cost of the next task drops to nearly zero. The human doesn't scale that way. The agent, theoretically, does.

The problem is that "theoretically" is doing a lot of heavy lifting in that sentence.

Omnicom's Quiet Power Play

OK so while everyone's debating whether agents are worth the money, Omnicom is already testing them to cut out ad tech middlemen entirely, according to Digiday. This is not a research project. This is a holding company with billions in media spend saying, out loud, that removing intermediaries is "an objective." If you work at a demand-side platform or a supply-side platform right now, that sentence should make your stomach drop.

What Omnicom is doing is essentially asking: what if the agent IS the ad tech stack? Instead of routing dollars through layers of platforms that each take a cut, you let an AI agent negotiate, place, and optimize directly. The middlemen have survived for years because the complexity of programmatic advertising made them necessary. Agents threaten to collapse that complexity into a single automated workflow. It's brutal and, honestly, kind of elegant.

The ad tech industry has been a rent-seeking paradise for a long time. Omnicom testing agents to bypass it isn't just a technology story. It's a power story. The holding companies have always resented the margin that intermediaries extract, and now they have a plausible weapon. Whether the agents actually perform better than the existing stack is almost secondary to the leverage this gives Omnicom in negotiations with every vendor in the chain.

Your Wallet, Their Playground

WIRED published a piece this week about the race to keep AI agents from running wild with your credit cards, and CBS News asked whether you should let agents do your shopping. These are two sides of the same coin, and the coin is your money. The consumer-facing agent, the one that books your flights and orders your groceries, needs access to payment methods. That's the whole point. But the moment you give an autonomous system spending authority, you've created a target that every scammer, every prompt injection artist, and every poorly tested API integration is going to probe.

Imagine ordering the same item twice because you forgot you had already clicked "buy." An agent can make that ordinary mistake across multiple accounts, so someone needs to check for duplicates before another payment goes through. The failure modes aren't dramatic cyberattacks (though those are possible too). They're mundane, expensive, and deeply annoying. And the guardrails? According to WIRED, we're still in the "race" phase. Meaning they don't really exist yet in any standardized way.

Forbes put it more bluntly: AI agents with no boss will go rogue. That framing is a bit theatrical (and yes, I know Forbes loves a dramatic headline), but the underlying point is sound. Autonomy without oversight isn't intelligence. It's chaos with a budget.

The Infrastructure Underneath

While the governance conversation spins its wheels, NVIDIA quietly dropped the Nemotron 3 Nano Omni model, which unifies vision, audio, and language processing and claims up to 9x more efficiency for AI agents, per the NVIDIA Blog. This matters more than it sounds. The bottleneck for agent deployment isn't just "can the model reason well enough." It's "can it run cheaply enough to justify running at all." A 9x efficiency gain on multimodal tasks means agents that can see, hear, and read simultaneously without melting your cloud budget. That's the kind of infrastructure improvement that turns demos into products.

Meanwhile, Nature published "Call your AI agent" on May 1st, a signal that the scientific establishment is taking the agent paradigm seriously enough to editorialize about it. And Andreessen Horowitz weighed in with a piece on asserting American leadership in open-source AI, which (reading between the lines) is really about making sure the models powering these agents aren't all locked behind proprietary APIs controlled by three companies in San Francisco.

The Real Bet

So here's what I think is actually happening this week, underneath all the headlines. The industry has collectively decided that agents are the form factor, the way AI gets deployed in the real world. Not chatbots. Not copilots. Agents. Autonomous, task-completing, money-spending agents. And every single story from the past seven days, from Omnicom's ad tech disruption to NVIDIA's efficiency chips to WIRED's credit card panic, is a consequence of that single decision rippling outward.

The part that nobody seems comfortable saying out loud is that we're deploying these things faster than we're learning to control them. Not in some sci-fi "Skynet" sense. In the boring, real sense of: who's liable when an agent overspends by $40,000 on a media buy? Who audits the agent's decisions after the fact? The companies building agents are moving at startup speed. The companies building oversight tools are moving at committee speed. That gap is where the interesting (and expensive) mistakes are going to happen, probably before the end of this year.