Personal Agents and Services

Working thesis

Personal agents may replace more app navigation than apps. A person increasingly expresses an outcome in one place, while many services still supply the records, capabilities, infrastructure, and responsibility needed to achieve it.

That shift changes the division of labor. The personal agent coordinates across a person’s life; services perform bounded work in particular domains. Memory should connect those domains without collapsing every record and conversation into one provider’s database. As this arrangement becomes technically workable, the central struggle is likely to concern economics: who controls access to the customer, who captures the value, and whose interests determine what gets recommended.

This is a working model, not a forecast that every product will become a background API. Reliability, security, authorization, and liability remain difficult. “The biggest conflict will be economic” is a long-run judgment about what may remain contested even after interoperability improves.

The central question is: what should remain valuable when a user no longer needs to visit a service’s interface?

1. A division of responsibility

A useful distinction begins with the principal each system serves.

A personal agent is intended to represent a person across domains. It can connect goals, preferences, schedules, budgets, and commitments that no single service sees in full. A trip involves work obligations, transport, accommodation, family plans, and personal trade-offs. Coordination across those concerns is the personal agent’s role.

A service supplies something more bounded: available inventory, an authoritative record, permission to perform a transaction, specialized computation, operational expertise, or a real-world outcome. A hotel can confirm a room and honor a reservation. A design service can maintain an editable project. A logistics provider can move a parcel. The personal agent’s ability to describe these outcomes does not create the capacity to deliver them.

“Personal” describes an intended relationship, however, not proof of alignment. A personal-agent provider may have commercial interests of its own. Likewise, a service’s agent may represent the service rather than the customer. Two agents can communicate perfectly while pursuing different objectives.

It is also useful to separate a service from its integration. A plugin is an interface to capabilities. It is not the inventory, operating organization, customer support, or business behind them. Replacing the interface does not necessarily replace the service.

Tools, delegated agents, and specialist interfaces

Different tasks deserve different boundaries.

A narrow operation with known semantics, such as retrieving a reservation or updating a field, usually belongs behind an API or tool. Adding a conversational agent does not automatically improve it. A task that requires local planning and adaptation may justify delegation to a domain agent, provided the goal, permissions, outputs, and stopping conditions are clear. Work involving dense visual judgment or continuous manipulation may still belong in a specialist interface.

The A2A documentation distinguishes communication with tools through MCP from collaboration between independently operated agents through A2A. It also describes collaboration without exposing internal memory or implementation. These are useful communication boundaries, not guarantees of competence, authorization, or aligned incentives.

Anthropic’s tool-design guidance emphasizes selecting useful tools, defining clear boundaries, returning meaningful context, and evaluating performance on realistic tasks. That supports a practical starting point: expose a well-scoped capability before assuming an autonomous counterpart is necessary.

The personal agent remains responsible for coordination and verification within its role. The service remains responsible for what it actually commits and delivers. Delegation should make those boundaries clearer, especially when work fails.

2. Agents may replace navigation more than applications

Much of app usage is coordination overhead: finding the right product, signing in, locating a function, re-entering context, comparing results, and carrying information to the next app. Personal agents could reduce that overhead without replacing the systems underneath.

Consider a presentation. The user can explain the audience and purpose to an agent, obtain a draft from a design service, and then open a visual editor for detailed work. OpenAI’s guidance on ChatGPT apps gives Canva as an example of this division: presentation drafting can happen within ChatGPT, while deeper slide editing remains in Canva’s editor.

The possible displacement is therefore uneven. An app’s home screen, menus, onboarding sequence, and routine forms may lose importance as entry points. Its project model, collaboration system, rendering engine, or execution capabilities may remain essential. Some interfaces may move into the agent experience as focused visual controls rather than disappear.

There are also activities where navigation is part of the value. Browsing can reveal preferences the user could not specify in advance. A visual canvas supports rapid exploration. A social space depends on shared participation. A professional dashboard can expose many simultaneous constraints more efficiently than a conversation.

The likely pattern is a mix of delegation and direct interaction. The agent handles context and transitions; the service supplies an interface when direct attention improves the result. The useful design question is where human interaction earns its cost, not how to eliminate every screen.

3. Memory should be connected, not indiscriminately centralized

Cross-domain assistance requires continuity. It does not follow that one agent provider should hold complete copies of every domain record, private conversation, and task history.

Three categories need different treatment.

Personal context includes preferences, priorities, relationships, and recurring constraints. Some of it should travel with the person across services. It should also remain inspectable and correctable: an inferred preference is weaker than an explicit instruction, and yesterday’s priority may no longer apply.

Domain records include reservations, orders, shared project files, account balances, and professional records. Their authority comes from the responsible system and its update process. An agent’s remembered summary can help find or interpret a record, but cannot establish that a booking still exists or a transaction settled.

Task state describes what is being done now: the agreed goal, granted permissions, completed actions, pending decisions, and evidence of completion. It must survive interruption without being confused with a lasting preference or an authoritative domain fact.

A connected memory architecture preserves these distinctions. It can retrieve the relevant record when needed, retain a reference and timestamp, and share only the context necessary for the next operation. Important claims carry provenance, freshness, and an indication of whether they were stated, observed, or inferred. Contradictions should trigger reconciliation rather than silent merging.

A hotel does not need a person’s entire travel conversation to answer whether an accessible room is available. A scoped request can carry dates and relevant requirements. OpenAI’s app-design guidance similarly recommends minimal structured inputs rather than sending whole conversations. The broader memory model here is a design proposal, not a claim that current integrations already implement it.

Centralization does offer convenience: faster retrieval, fewer dependencies, and easier cross-domain synthesis. Selective caches and user-controlled stores can be useful. The objection is to indiscriminate copying and ambiguous authority, not to every central component.

Connection also needs meaningful controls. Users should be able to inspect what was shared, revoke future access, and move useful personal context to another agent. Revocation cannot magically erase copies already disclosed; retention and deletion need separate treatment. Portability must preserve sources, uncertainty, and permissions, not merely export a pile of text.

These are economic choices as well as privacy choices. An agent that understands a person deeply but makes that understanding difficult to move can become costly to leave.

4. The enduring conflict is likely to be economic

Imagine flawless technical interoperability. Every service can be discovered, every permission is correctly scoped, and every task returns a reliable result. A hotel recommendation still raises a conflict.

The traveler wants the best fit for the trip, including total cost, location, comfort, and flexibility. A hotel wants profitable occupancy and may favor direct bookings. A marketplace may favor transactions that generate higher commissions. An agent provider may earn subscriptions, referral fees, advertising revenue, or returns from its own services.

Those objectives sometimes overlap. They do not always converge. Better protocols make exchange possible; they do not decide whose objective governs it.

Who controls the customer relationship?

The customer relationship is a bundle of roles, not a single asset.

One organization may receive the initial intent. Another may determine the shortlist. A third may process the transaction, while a fourth delivers the experience and handles failure. The user may trust each for a different reason.

Personal agents could gain influence by becoming the place where intent is first expressed. A service excluded from the shortlist may never get the opportunity to compete, regardless of how polished its own app is. Control over discovery and routing can therefore become a powerful gatekeeping position.

Services can retain influence through distinctive supply, reputation, loyalty, direct relationships, or capabilities the agent cannot substitute. They may also restrict access, offer different terms through different channels, or insist that parts of the transaction happen in their own environment.

The important questions become concrete: who gets to explain the choice, who can contact the customer afterward, who handles a dispute, and who can be replaced without losing continuity? A friendly conversational interface does not settle any of them.

Who captures the value?

An agent can create substantial value by reducing search and coordination costs. The service creates value by delivering the outcome. The distribution of payment between them depends on scarcity, alternatives, and bargaining power.

If several services produce genuinely interchangeable results, an agent can compare and switch among them easily. Competition may push prices down and transfer some benefit to the user or the agent platform.

If a service has scarce supply, trusted shared records, superior execution, or a result that is hard to reproduce, its bargaining position may strengthen. Agents could make that capability easier to discover and use, expanding demand without removing its value.

Value creation and value capture can therefore move in different directions. A service may receive more usage while losing margin to an intermediary. An agent may be extremely useful yet struggle to charge enough to cover its costs. A platform may capture fees because it controls distribution rather than because it contributes the hardest part of the work.

The business model matters. Subscriptions, usage charges, referral payments, and bundled services distribute incentives differently. No model eliminates conflicts automatically.

Whose interests determine the recommendation?

Recommendation quality includes both factual accuracy and the objective being optimized. An accurate description of a hotel can still support a ranking that favors the wrong interest.

A user-aligned agent should make the relevant trade-offs visible: why an option fits, which inventory was searched, what alternatives were excluded, and whether a commercial relationship affected the recommendation. It should distinguish a preferred option from a sponsored placement and disclose when it cannot compare the full market.

That is a normative standard, not a prediction that the market will reliably provide it. Paying a subscription may reduce reliance on advertising, but does not remove incentives to retain users, promote affiliated products, or lower service costs. An open protocol likewise does not guarantee neutral ranking, portable memory, or fair access.

The deepest question is whether the agent is an accountable representative of the user or a more persuasive distribution channel. The same interface can support either role.

5. Some services become commodities; others become more valuable

Agent-mediated access makes it easier to compare capabilities separately from their interfaces. That could expose products whose differentiation depends mainly on familiar screens, routine data entry, or generic reasoning that an agent can reproduce with comparable quality.

A thin interface around an interchangeable capability is vulnerable. A narrow feature may be absorbed into the agent or become one supplier among many. But describing a product as “CRUD” or “a wrapper” is not enough to predict its fate. Apparently simple software can coordinate multiple people, enforce important rules, preserve trusted records, and support difficult operational work.

More durable advantages may come from:

• Hard-to-reproduce observations: measurements or evidence that competitors cannot simply generate from a prompt.
• Authoritative shared state: records whose usefulness depends on agreement, permissions, continuity, and a reliable update process.
• Execution capacity and scarce supply: the ability to deliver something in the world, not merely recommend it.
• Operational expertise and validated outcomes: repeatable performance under messy conditions, including exception handling.
• Accountability: a credible party that stands behind the result, supports recovery, and accepts defined responsibility.

These advantages still have to be demonstrated. Possessing data does not ensure that it is unique, lawful to use, current, or valuable. Domain prompts alone may be easy to copy. Claimed expertise matters less than measured outcomes. A service cannot assume that a horizontal agent company will remain outside its market.

The direction of change may nevertheless be positive for strong specialists. If an agent removes discovery and integration friction, a valuable service can reach customers who would never have installed its app. Less visible branding can coexist with greater underlying demand.

A useful distinction is between losing screen time and losing indispensability. They are not the same business event.

6. What could weaken this thesis?

The proposed division of labor depends on several conditions.

First, users must trust agents enough to delegate meaningful work. If failures remain expensive, hard to detect, or difficult to reverse, direct interaction may remain the preferred route. Technical reliability can be the immediate bottleneck even if economics is the longer-run conflict.

Second, services must allow useful access. If important capabilities stay behind exclusive arrangements or restrictive interfaces, the ecosystem may fragment. Open standards can reduce integration costs without removing a provider’s incentive to withhold access.

Third, a personal agent must offer better coordination than a well-designed specialized product. In some domains, the specialist already understands the task, presents the relevant state clearly, and supports expert judgment. Adding a conversational intermediary can increase latency and ambiguity.

Finally, market structure is not predetermined. A few agent platforms might dominate distribution, several interoperable assistants might compete, or domain-specific agents might preserve direct customer relationships. Agent companies may integrate vertically; service companies may build their own general entry points.

These alternatives suggest a more limited claim: agents are likely to change where coordination happens. How far that shifts bargaining power, and in whose favor, remains an empirical question.

7. A practical test for builders

The most useful test is:

If an excellent personal agent replaced our user interface, what would it still need from us, and why would it choose us?

A strong answer names an outcome, an asset, or a responsibility that remains valuable without requiring the user to navigate the product. It also explains how another system can verify that value.

From there, the product work becomes clearer: expose a few reliable capabilities, return structured evidence and durable references, preserve authoritative domain state, and keep specialist interfaces where they improve decisions. Support bounded delegation without demanding the user’s complete memory.

The commercial work is equally important. Decide whether the business can thrive when someone else controls the entry point. Understand which incentives might distort recommendations, what customers can take with them when they leave, and what earns a direct relationship beyond habit or friction.

The strongest position may be to become easier for a good personal agent to use while remaining difficult to replace for a good reason.

Evidence, connections, and open questions

The sources establish narrower technical and product observations than the economic thesis:

• OpenAI: What makes a great ChatGPT app grounds the examples of focused capabilities, specialist interfaces, and limited data sharing. It does not establish how future profits will be distributed.
• A2A Protocol grounds the distinction between tool access and independent-agent collaboration. Interoperability alone supplies no guarantee of alignment.
• Anthropic: Writing effective tools for agents grounds the emphasis on tool scope, context, and evaluation.

This topic connects engineering ideas about control, state, verification, and interoperability with product strategy and political economy.

The current view should change with evidence. Particularly useful questions are:

1. Does agent-mediated use reduce direct app visits while preserving or increasing service usage and revenue?
2. When agents can compare suppliers reliably, do savings reach users or become new intermediary fees?
3. Can users move meaningful memory between agents without losing provenance, permissions, or important context?
4. How often do commercial relationships alter recommendations, and can users detect or override that influence?
5. Which services gain demand because agents reveal their value, and which survive mainly through restricted access?
6. When something goes wrong across several providers, can the user identify who is responsible and obtain a remedy?

The emerging relationship should be judged by those outcomes, not by the number of agents, integrations, or conversations it produces.