Two Gartner numbers from the last year only make sense together. The first: up to 40% of enterprise applications will ship task-specific AI agents by the end of 2026, from under 5% in 2025. The second, from the same firm: of the thousands of vendors selling "agentic AI," roughly 130 are the real thing. The rest are rebadged chatbots, assistants, and RPA scripts. Gartner has a word for the practice — agent-washing — and a forecast to match: more than 40% of agentic AI projects will be canceled by the end of 2027.
You cannot reconcile "40% adoption" with "130 real vendors" as facts about the same technology. You can only reconcile them as facts about a word. "Agent" stopped being a technical claim and became a marketing one. It now means roughly what "AI-powered" meant in 2019: present on the box, absent from the product.
This matters beyond pedantry, because companies are buying the word. A budget line gets approved for "an agent," a wrapper gets shipped against it, and eighteen months later the project is in the 40% that gets canceled. The thesis of this post is simple and expensive to ignore: you cannot buy a capability by buying its name, and the gap between the name and the capability has a P&L.
The word ate itself
The claims are running years ahead of the systems.
Writer's 2026 survey found 97% of executives saying their company deployed AI agents in the past year. Zapier put deployment-or-testing at 72%. When almost everyone reports the same thing, the term has stopped carrying information. A word that every product satisfies is not a specification. It is a mood.
Agent-washing is not usually a lie. It is a stretch. A support chatbot that can look up an order status gets called an agent. A macro that files an expense report gets called an agent. A prompt template with a nice UI gets called an agent. Each one does something, and "agent" is vague enough to cover it, so the marketing team reaches for the word that wins the meeting. The vendor is not inventing a capability out of nothing. They are relabeling a real but smaller one with a bigger name, and the buyer has no cheap way to tell the difference from the outside.
That is the trap. The label is free. The capability is not. And the two have come completely unbundled in the market.
A wrapper answers. An agent acts.
There is a clean line under the marketing, and it has nothing to do with which model you use.
A wrapper is a straight line: prompt in, model call, text out. It can be brilliant text. It can use the best model on the market. It is still a wrapper, because it does two things a person does not get paid for — it produces language and it stops. It takes no action against a system of record, and it has no way to know whether what it produced was right.
An agent is a loop. It holds a goal, plans a step, takes an action that changes state — books the slot, moves the ticket, updates the record, sends the message — then observes the result and evaluates whether the action moved it toward the goal. If not, it tries again. Two properties make it an agent, and a wrapper has neither: it acts, so there is a real-world consequence someone owns, and it checks, so it knows when it failed.
The eval is the part that gets skipped, and skipping it is the tell. A system that acts but cannot evaluate its own action is not an agent that is a little unreliable. It is a wrapper with permissions, which is worse than either, because it produces consequences it cannot assess. Most of what fails in production fails here: not because the model was weak, but because nothing in the loop knew the difference between a good action and a bad one.
Why the fake ones ship anyway
If the line is that clean, why is the market full of wrappers wearing the agent label? Because the incentives point that way at every step.
A wrapper is cheap to build and cheap to run. It makes one model call and stops, so it never pays the token bill a real agent pays. I wrote a few weeks ago that an agentic task burns 5 to 30 times the tokens of a chatbot, because the loop runs, retries, and checks itself. Turn that around and the wrapper's economics are obvious: skip the loop and you skip the cost. Agent-washing is, in part, cost avoidance dressed as a product decision.
A wrapper also demos beautifully. On stage, one clean prompt and one clean answer is exactly what an agent looks like from the outside. The loop, the retries, the evals, the boring machinery that separates the two never appears in a two-minute demo. So the wrapper wins the pilot, wins the budget, and then meets production — where the actions have consequences and there is no eval to catch the bad ones. That is where Gartner's three cancellation reasons come from. Escalating cost, unclear value, and weak risk controls are not three separate failures. They are what it looks like when a label ships without the loop underneath it.
The two questions I ask before calling anything an agent
Because "is it an agent?" is now unanswerable — every vendor says yes — I stopped asking it. Two better questions do the sorting, and they work from the buyer's side of the table without reading a line of the vendor's code.
Does it take an action I would otherwise pay a person to take? Not "does it generate a draft." A draft is language, and language is the wrapper's whole job. The question is whether it commits something: a booking, a status change, a record written, a message sent to a real counterparty. If a human still has to take the consequential action afterward, you bought a very good assistant and you should price it as one.
Does it know when it failed? This is the eval, and it is the question almost no washed agent survives. Ask what signal tells the system its last action was wrong, and what it does next when it gets that signal. If the answer is "a human notices," there is no loop. There is a wrapper with write access and a person doing the checking the system was supposed to do.
Two yeses and you have an agent worth the token bill. One or zero and you have something useful with the wrong name and the wrong price tag.
What this looks like at Shakers
A marketplace is a good place to see the line, because a wrapper here is obvious and cheap and an agent has to earn its cost.
Drafting an outreach message to a freelancer is a wrapper. It is language, a human sends it, and nothing changes until they do. Useful, but not an agent, and we do not call it one. The part of Matchmaking that behaves like an agent is the part that takes an action against the match itself and then checks whether the action worked: it ranks and re-ranks candidates against a live role, and an eval scores whether the surfaced set actually converts toward a started project. When the eval says the action did not improve the match, the loop runs again or the step does not ship at all. The consequence is real, and something other than a person knows whether it was right.
The general rule, for anyone running this in production: if a step cannot be evaluated, it does not get shipped as an agent. It ships as a suggestion a human owns, priced and described as exactly that. Naming it honestly is not modesty. It is what keeps the thing off the 2027 cancellation list, because you never sold a loop you had not built.
Closing
Agent-washing looks like a vocabulary problem and gets treated as one, which is why it keeps working. It is actually a capital allocation problem wearing a vocabulary problem's clothes. Money is being committed against a word whose meaning collapsed, and the reckoning is already scheduled: Gartner's 40% cancellation forecast is the market repricing the word by 2027, project by dead project.
You do not need to wait for that. The line was always there under the marketing. An agent acts against a system you care about, and it knows when it was wrong. Everything else is a wrapper, and a wrapper is a fine thing to build and a terrible thing to buy under another name. Ask the two questions. Build the loop or do not call it an agent.