Who gets paid when agents work?
5 min read
Meta launched Muse, and Zuckerberg is already talking about where the money will come from.
In his September 8 conversation with Alex Heath, he described a personal agent with a large free allowance, subscriptions for heavier use, and eventually a small cut of transactions, potentially paid by the businesses involved. An agent could help you save money or earn it, and Meta would get paid along the way.
Meta says Muse keeps conversations and VM data out of its ad systems. It could still earn money by helping decide where people spend theirs.
Muse, Grok Bot, and Instinct are competing for the personal-agent role. Hermes gives people a more self-managed option. Codex, Claude Code, and Devin focus on software development; Rogo on finance; Harvey on legal work. What interests me here is how much of a decision someone is willing to hand over.
Filling out your address leaves the purchase decision with you. Letting an agent choose which groceries arrive at your door hands over much more.
Stripe's 2025 annual letter reports $1.9 trillion in business volume for 2025, up 34%, and more than 200 million Link users. These describe its existing payments business, not agent-generated sales.
The letter sets out five levels of agentic commerce. Using groceries as an example:
Eliminating web forms
YouHere's my basket. Order it for delivery tomorrow.
The agent fills in your details and pays. You already made the choices.
Descriptive search
YouFind groceries for five quick vegetarian dinners for two.
The agent finds suitable options. You choose what goes in the basket.
Persistence
YouFind groceries for next week's dinners.
It remembers the meals you like and your budget. You still choose what to buy.
Delegation
YouOrder next week's groceries. Keep it under $100.
The agent chooses the basket and pays within the limit you set.
Anticipation
Your agentNext week's groceries are ordered. Delivery is Sunday.
It notices what you need and orders under your previously agreed preferences and spending limits.
When the letter appeared in February 2026, Stripe placed the industry near the first two levels. The later stages describe a possible progression, not measured adoption.
At the first stage, the supermarket and the products have already won your attention. The agent saves you some typing. By the second, it helps decide which options you see. By the fourth, it can allocate your grocery budget without you comparing every brand and retailer.
A supermarket could have a beautiful app and never get opened. A smaller retailer could win an order because it has the right stock and can deliver on time. Or the agent could keep returning to the same large supplier because that relationship is easier to use. Making shopping more convenient does not tell us which of those outcomes will follow.
At the fifth stage, you no longer have to remember to start the purchase. For something recurring, like coffee or groceries, that sounds useful. I would still want to set the boundaries beforehand. Knowing what I usually buy would not be permission to spend any amount or replace it with whatever happens to be available.
The same change will show up in businesses. An employee might ask an agent to renew a software license within a company policy. A procurement agent might compare vendors, but still need an approved budget and someone accountable for the purchase. The product has to enforce limits on what the agent may choose, for whom, and at what price.
Muse already supports checkout through Stripe's Link. Zuckerberg's proposed transaction business is still a plan, with no disclosed commission schedule. But the reason to make the agent widely available is easier to understand if it can eventually participate in purchases people already make every week.
Instinct's Noah Shinn describes a similar ambition. He wants the assistant to remain free for users, with merchants paying a percentage of purchases in exchange for the business it brings them. His September conversation with Patrick O'Shaughnessy describes the proposal without establishing whether Instinct already earns those fees.
Free access does not mean the model is free to run. Someone pays for the compute and the work that fails. Transaction revenue could help cover an agent's cost, including the tasks that never produce a purchase. That would let its provider compete differently from a company that needs every user to pay a subscription.
The person using the agent wants a good purchase. A business wants the right item under its policy. The retailer wants an order, and the agent's provider may want a fee. Those interests can agree, but I would not assume they always will.
Stripe has a possible route across that competition. It can process purchases from several agents without needing everyone to choose the same assistant. A retailer will still need to receive the payment and handle a return, even if the customer never sees its checkout page. I would expect payment companies to compete for that work as more buying starts inside conversations.
My guess is that generic paid assistants will face pressure from products supported by transactions or another existing business. The same person could use a free agent for everyday requests and still pay for a service it cannot deliver well enough. A business could make that choice too. Payment providers can serve both the general agent and the specialist. There may be several winners in the same purchase, collecting money for different parts of it.
Before letting an agent handle my groceries every week, I would want to know what happens when the best deal comes from a supermarket that does not pay its provider. I would also want it to recommend skipping the order when I already have enough food, even if that means no transaction fee.