In the last post we argued that agents will come in two kinds: a personal agent represents one person, and a service agent represents a service provider. The deepest difference between them is the context each one accumulates. This post follows one consequence. If a service provider doesn't build its own service agent and only exposes an API for other agents to call, its information will be gathered and retold by the buyer's agent. And the buyer's agent will not speak for it.
The cost is concrete. The provider gets compressed into a few comparable fields, loses the chance to make its own case, and loses the chance to learn from every interaction.
Consumers will see information that has already been synthesized
Today, someone looking for a course opens a few schools' websites, reads the descriptions and reviews, asks support a question or two, and decides. Soon they're more likely to tell their personal agent "find me a course that fits," and let it search, compare and ask, then come back with three options and a recommendation.
What the consumer receives is no longer each provider's own information. It's the agent's synthesis. Which providers get mentioned, how they're described and in what order are all the agent's call.
None of this is new. It's what a buyer's agent in real estate does: tour ten houses, come back and talk about three. AI agents just make this kind of representation cheap enough for everyone to use for everything.
The buyer's agent is loyal to the buyer
Last time we described the principal–agent relationship: an agent acts for a principal and owes that principal loyalty. A personal agent's principal is the consumer. When it synthesizes providers' information, its job is to make the best choice for the consumer, not to get any particular provider chosen.
That's exactly what it should do. For the provider, though, the implication is blunt: in that conversation, nobody represents you.
No home seller works this way. Nobody skips the listing agent, hands the house's details to the buyer's agent and hopes the buyer's agent pitches it well. The buyer's agent can be honest and skilled, but the job is to push the price down and find the flaws. The seller's interests have to be argued by the seller's own agent.
A provider that only offers tools is doing exactly this.
Only the most comparable fields survive synthesis
A tool hands over data: price, availability, opening hours, specs. These fields are uniform and easy to compare, so they're also the easiest to fold into a recommendation.
What really separates one provider from another usually isn't in those fields. Is this teacher good with complete beginners? Has this lawyer handled cases like yours? Can this technician find the real fault? This is information the provider knows and outsiders find hard to verify.
George Akerlof analyzed this situation in his 1970 paper "The Market for 'Lemons'." When buyers can't tell quality apart, they pay for average quality. Good sellers don't get the price they deserve and drift out, and competition collapses onto price. If information about service quality can't get across, comparison lands on price.
Personal agents will speed this up. They're good at comparing structured fields, and whatever you can't state and they can't ask about doesn't exist in their recommendation. A provider that only offers tools gets squeezed, bit by bit, into an interchangeable commodity.
Tools also hand your context to the other side
Being compared on price is only the surface. The deeper loss is context.
Every interaction produces context: what the customer asked, what they cared about, why they didn't buy, what disappointed them afterward. Today that context shows up in a provider's support chats, consultation notes and after-sales tickets, where the provider can learn from it and improve.
When the conversation happens between a personal agent and a tool, that context settles on the personal agent's side. The agent gets to know this customer better and better, and it learns the strengths and weaknesses of providers like you. What you see is a stream of calls: one price lookup, one availability check. You don't know why the customer left, so you can't improve.
As we said last time, a service agent evolves version by version by distilling what it learns from the cases it serves. Without context there are no cases to distill. A provider that only offers tools never grows a service structure of its own.
Providers need a context system of their own
So what a service provider needs isn't a better API. It needs a context system it owns, one that meets three conditions.
Buildable. The provider can turn its business knowledge, service rules and professional judgment into a structure an agent can execute, instead of handing over a few data tables.
Accumulating. The context from every interaction stays with the provider and builds up as its own asset, instead of being scattered across someone else's conversations.
Iterable. The provider can find problems in the accumulated cases, change the structure and ship a new version, so the service keeps getting better.
The service agent is what carries this system. It's the provider's own agent: it speaks for you in the conversation and makes your judgment and capabilities visible; it accumulates for you, so context stays on your side; and it iterates by your rules and answers to your customers.
This is the premise for service agents existing at all. A service agent isn't a chat widget bolted onto a website. It's how a service provider keeps a representative of its own once agents do the errands.
That doesn't mean turning personal agents away. Consumers' agents will still come to you, and you should serve them well. The difference is who meets them at your door: your own service agent, not an endpoint that only returns data. Deals often go more smoothly when buyer and seller each have an agent. How the two kinds of agents meet and work together is the subject of the third part of this series.
Products can be compared; services need their own agent
This argument has limits, and the limit sits in the word "service."
In 1970 the economist Phillip Nelson split goods into two kinds. Search goods can be judged before you buy, like a TV's size, resolution and price. Experience goods can only be judged after you use them. In 1973 Michael Darby and Edi Karni added a third kind, credence goods, which are hard to judge even afterward, like a surgery or a piece of legal advice.
Products are mostly static. The same model has the same spec wherever you buy it, and its quality is written in the specs. Letting personal agents compare these is a good fit: accurate, complete product data that agents can find easily may be all you need. Being compared on price may suit you fine.
Services are dynamic. They vary by person: the same course is taught differently to a beginner and to someone with experience. They're produced together with the customer: a lawyer has to understand your situation before giving advice. Their quality often shows only over time, and sometimes never clearly. Most services are experience or credence goods, exactly the kind that resists being squeezed into a few fields.
So the more something is a service, the more the provider needs someone to represent it, explain it, deliver it and answer for the result. Education, healthcare, law, repairs, financial advice and public services all work this way. That's why we call it a service agent: what it represents isn't a product with a uniform spec, but a service that takes judgment, varies from person to person and carries consequences.
Next
If service providers are to build context systems of their own, the next question is where that system builds up, and why the service agent that carries it is itself an asset. That's the next post: "Service agents are a new form of asset."