Trick or Treat? Agent Sprawl Is Already at Your Door

Halloween runs on one rule. Porch light on, you knock. Porch light off, you keep walking.

No contract, no committee, no badge reader. A whole neighborhood of strangers runs on a single bulb, and it works.

Most health systems let AI agents in with no porch light at all. They knock, somebody opens the door, and they walk straight to the kitchen.

How the neighborhood filled up

In a recent Dreamforce interview, Chamath Palihapitiya gave this problem a name: agentic sprawl. His timeline maps almost perfectly onto what we see in healthcare.

Wave 3 is where the porch lights went out. For health plans, we've called it buying a hat on a hat, a new AI layer bolted onto every point solution already in the stack.

Same model, different costume

Here's the thing about trick-or-treaters. Take the masks off and a lot of them are the same few kids on their second lap.

Agents work the same way. Peel the vendor wrapper off your prior auth agent, your denials agent, and your ambient scribe, and you'll often find the same handful of frontier models underneath.

Chamath made the sharper point. Model and harness come apart. You can run one lab's harness on another lab's model, or put an open model inside either, and the price and precision curves change with every pairing.

That's the lever. If the model is swappable, the harness is the asset. Own it and you decide which model runs which job, at what cost, under which policy. Rent a different one from every vendor and those decisions belong to someone else. (We made the long version of this case in The Model Gets the Headlines. The Harness Does the Work.)

Check the candy before you eat it

Every parent inspects the haul before bedtime. We built χ-BENCH to do the same thing for agents.

We ran 30 frontier agents through 75 long-horizon workflows in prior authorization, utilization management, and care management, inside a simulator of 21 healthcare apps. Each trial runs 60 to 80 steps with no do-overs once a step commits.

Source: χ-BENCH, ACTAVA with Johns Hopkins Medicine, Wellstar Health System, and academic collaborators, May 2026.

Those numbers describe one agent at a time. Sprawl stacks them.

Every handoff between two vendors' agents is another end-to-end run, and end-to-end is exactly where the benchmark hit zero. Ten ungoverned agents passing work to each other is a haunted house with ten doors and no floor plan.

The knock you don't want

On October 31, the knock is a kid in a cape. The other 364 days, it's an auditor, a regulator, or a plaintiff's attorney with one question. Why did this happen?

Chamath's point here landed hard. You can't walk down the hall and ask the agent. It has no first and last name, and a judge won't accept a shrug.

Healthcare figured this out long ago. "The agent acted on its own" has never been a defense, and it won't become one because the agent came from a vendor.

"Accountability requires knowing which agent did what, under which policy, and with what result." Dr. T. Y. Alvin Liu, Director, Gills AI Innovation Center, Johns Hopkins Medicine, to Healthcare IT News

Every agent in your building should answer all three. Most sprawl can't answer the first.

The "Take One" bowl

Every block has one house. The owner leaves a bowl on the step with a sign that says Take One, then goes inside to watch a movie.

You know what that bowl looks like at 7:15.

Chamath called the enterprise version token maxing, agents running with no guardrails and burning compute because nothing tells them to stop. An agent with no owner, no budget, and no ceiling is that bowl. It gets emptied, and the invoice arrives in November.

We've argued for a while that you should budget agents like capacity, with a cost per unit of work and a hard ceiling. Johns Hopkins tracks token consumption and per-agent cost against baselines, and holds off on reporting financial return until the evidence earns it. That's what an agent earning its place looks like.

Porch-light rules for a digital workforce

Chamath's diagnosis is that the org chart breaks once hundreds of agents run long tasks on their own. Hierarchy means nothing unless someone writes it down and enforces it with guardrails, tripwires, and checkers checking the checkers. His estimate of how much of that infrastructure exists today was zero.

We agree, and we'd push one step further. The org chart stays. It gets a second column for Non-Human Resources, where every agent gets hired, onboarded, reviewed, and retired like any other worker.

Here's the porch-light test we'd run on every agent before it gets through the door.

The third row is where the checkers come in. An agent that passed in March drifts by October when a payer rewrites a policy. AgentOps is the discipline of evaluating, remediating, and re-certifying on a schedule, with a second system checking the first.

Your staff has seen the horror movies

Chamath said change management is the hardest part of his enterprise work, harder than the technology and harder than executive buy-in. His teams spend much of their time undoing the doom story people absorb outside the office.

Healthcare staff have seen the same movies. A UM nurse has read the headlines about algorithms and denied care, and she has every reason to side-eye the next agent in her queue.

Sprawl makes her right. Agents show up unannounced, from vendors she's never met, doing work she can't see.

Governance is how you earn her back. Show her the owner, the policy, the test results, and the exact step where she signs off. Human-in-the-loop is the steering wheel, and people ride more calmly when they can see whose hands are on it.

Where ACTAVA sits

ACTAVA is the AI factory for healthcare. Master your agentic future. We give every agent in your enterprise a porch light, whichever vendor built it and whichever model runs underneath.

Happy Halloween. Leave the porch light on for the agents you actually invited.