Follow-up: Four kinds of agents: general, service, personal and assume adds the two other kinds from the product side.
When people talk about AI agents today, they almost always mean personal agents: an assistant that answers your email, books your trips and runs your calendar. That is only half of the picture. Human society has always had two kinds of agents: one that works for you, and one that serves you on behalf of a provider. AI will rebuild both, so the other half, the service agent, is coming too.
This is worth getting right now. A provider that turns itself into nothing more than an interface for other people's agents will slowly lose its relationship with its own customers, until all that is left to compare is price and stock.
Two kinds of agents: personal agents and service agents
We use two names to tell them apart.
Personal Agent: an agent that represents one person and answers only to that person. It follows that person's life over time, builds up private context about them, and does as much as it can for them within the bounds of law and ethics.
Service Agent: an agent that represents a service provider and serves many people on its behalf. It is a stable service structure that, within the scope the provider grants it, applies the same rules to everyone it serves.
| Personal Agent | Service Agent | |
|---|---|---|
| Represents | One person | A service provider: a business, a government office, a school, a hospital, or an independent lawyer |
| Nature | An ongoing process | A stable structure |
| Seeks | Change: it keeps state and keeps growing | Constancy: it stays stateless and stable |
| Grows | Vertically: deeper into one person's life | Horizontally: the same structure serves more people |
| In a transaction | The buyer's agent | The seller's agent |
A service agent is more than customer support. Support is one kind of service; after-sales, bookings, shopping help, applications and tutoring are others. Nor is it limited to business: a public service counter, a hospital triage desk and a school registrar all do the same kind of work.
The name matches the term 服务智能体 (service agent) in China's national standard GB/Z 185—2026 on agent interconnection. The standard divides agents by their role in a single interaction: the side that provides a service is the service agent, the side that asks for it is the requesting agent. We divide them by whom they represent. In practice the two line up: in an interaction, the service agent plays the providing side and the personal agent usually plays the requesting side.
Society has always had these two kinds of agents
Economics and law have studied acting on someone else's behalf for a long time under the name of the principal–agent relationship: an agent acts for a principal and owes that principal loyalty, confidentiality, obedience and care.
The easiest place to see the two kinds is buying a house. The buyer hires a buyer's agent to find homes, compare prices, negotiate, and keep the budget and walk-away price confidential. The seller hires a listing agent to market the house and get a good price. Both are professionals and both must be honest, but they are loyal to different people.
Take away the house and you get two broader kinds of agents. One acts for a person across the many things in their life and work. The other acts for a provider in front of all its customers. The first plays the buyer's agent when it shops; the second plays the seller's agent when it faces a buyer.
Personal agents are the new form of the first kind, and service agents are the new form of the second. AI changes what agents can do and what they cost. It does not change the structure of agency itself.
One agent cannot sit on both sides
If both kinds will be rebuilt, could a single agent serve both sides? Economics says no, for three reasons.
First, splitting the gains from a deal is zero-sum. A deal happens because the most a buyer will pay is above the least a seller will accept, and both sides create the gap between them. Closing the deal is good for both, but every bit of that gap given to one side is taken from the other. The buyer's agent pushes the price down and the seller's agent pushes it up. One agent cannot push both ways.
Second, an agent is valuable because it keeps its principal's private information. The buyer's budget, the seller's floor, a product's flaws: these are things the agent knows and the other side does not. An agent that holds both sides' cards will sooner or later betray one of them. That is why, in real estate, one agent representing both buyer and seller is usually tightly restricted and must be disclosed to both.
Third, an agent is loyal to whoever pays it and can fire it. An agent that claims to work for users while earning its money from someone else lives with a permanent conflict of interest. Ad-funded search has faced exactly this question for years.
All three reasons are structural. Making agents more capable does not remove them; it raises the stakes, because a smarter agent working both sides is better at betraying one of them.
One seeks change, the other seeks constancy
The conflict of interest explains why the two cannot merge. It does not yet explain why they grow into such different things. The deeper reason is that they want opposite things.
A personal agent seeks change. It serves one person's ongoing life: the hotel it booked for you today shapes the flight it picks tomorrow, and it still remembers the preference you mentioned last month. Its value comes from remembering you, understanding you and changing with you. So it needs state, long-term memory, and context that keeps growing.
A service agent seeks constancy. It serves through a structure: the same question gets the same rules no matter who asks, and the same application goes through the same process no matter who submits it. Its value comes from being stable, predictable and fair to everyone. So it needs to be stateless, starting every interaction from the same structure.
You might object that good service remembers regulars, and that a support agent can pull up your order history. Isn't that state? There are two layers here. A service agent can read data about you, but that data is the provider's business record, an input to its rules. It does not change the rules because of you. Another customer in the same situation gets the same treatment.
So the precise version is this: a service agent is constant across people, and a personal agent changes with its person. The moment an agent bends its rules for one person, it is no longer providing a uniform service; the moment it treats everyone the same, it is no longer anyone's own agent. Constancy and change cannot both be one agent's goal, and that is why the two kinds of agents must split.
That makes them different in every way that matters
From this opposition, the two kinds of agents move in opposite directions along every dimension.
They accumulate different things. A personal agent accumulates state: context about one person, with every experience recorded as it happens. A service agent accumulates structure: what it learns from the cases it handles does not quietly stay in one conversation. It is distilled, reviewed and released as a new version of the structure. The structure evolves, but every version is stable.
Their economics differ. A service agent runs on economies of scale: building the structure costs a lot up front, but serving one more person costs almost nothing, so it earns by growing wide. A personal agent accumulates what economists call a specific asset: context about one person is valuable only for that person and useless for anyone else, so it earns by growing deep. One is a general asset that pays through reuse; the other is a specific asset that pays through depth.
Their accountability differs. The consequences of a personal agent's actions fall mostly on its user, so it can go as far as the user asks within law and ethics. Every word and action of a service agent is the provider's responsibility toward all of its customers, so the provider draws its scope precisely, refuses requests outside it, and puts "make no mistakes" ahead of "satisfy every request". And only a stable structure can be tested and verified: something that changes every time cannot be guaranteed to be right for everyone.
These differences already show up in engineering. On our platform, a personal agent is bound to a persistent computer that remembers everything it has done, while a service agent gets a clean environment for every session, so each interaction starts from the same point.
Service agents will be a category of their own
If the two kinds differ in what they accumulate, how they make money and who answers for them, the service agent will not be a feature of personal agents. It will be a category of its own.
The natural position for companies building personal agents is the demand side: owning the user's entry point and choosing among providers on the user's behalf. They have neither the reason nor the standing to build a stable service structure for any one provider. And what providers need, personal agents cannot give them: turning their own knowledge and rules into an executable structure, putting it into their own website, app and channels, controlling what it may and may not do, keeping a record of every interaction, and standing behind the result.
Real estate split this way long ago: some agents list homes for sellers, others find homes for buyers, and they are different businesses. We expect agents to follow the same division.
The two will meet on a network
Splitting is not the same as fighting. Buyers' and sellers' interests are partly opposed, which is why their agents must be separate. They are also partly aligned, since both want the deal done and the job finished, which is why their agents will meet, work together and close.
Real estate solved this with a shared listing system: listing agents publish homes, buyers' agents search them, and the two find each other there before negotiating, each for its own principal. Agents need the same kind of infrastructure: service agents publish the services they offer, personal agents discover them and talk to them, and each gets the job done within its own bounds. The industry has started building it.
So the future is not personal agents swallowing everything. It is a network of both kinds. This series will follow that thread: the next post explains why service providers need service agents of their own, instead of being just an interface for someone else's agent.