Can you change agents without starting over?

Access shapes the usable shortlist

An agent can correctly compare the offers it can verify while excluding useful offers outside its permitted or supported routes.

Atlas’s hotel journey illustrates this mechanism; no named service or model was tested.

Convenience and commercial incentives can diverge

Delegation may remove opportunities to steer customers; providers therefore have incentives to shape access and intermediation.

Attributed operator and economic perspectives, not a finding that every provider restricts access or biases rankings.

AI-summary visits had fewer traditional-result clicks

Pew observed traditional-result clicks on 8% of Google visits with an AI summary and 15% without in March 2025.

900 US adults; observational query groups. Not causal, global or a measure of autonomous purchases.

Information needs a viable route to funding

If agent-mediated answers reduce valuable visits, publishers may seek paid access or restrict access; these responses can affect what agents can read.

Atlas’s conditional inference; Cloudflare’s proposal describes a commercial direction, not proven publisher economics.

Identity, authorization and order completion are distinct

The August 2026 UCP design separates platform, business and credential-provider roles; the business completes the transaction using a token passed by the platform.

A documented design with optional AP2 mandates, not evidence of hotel adoption, universal implementation or transaction reliability.

Switching requires more than exporting a conversation

Useful context, records, service access and permissions determine whether another agent can continue a task.

CMA identifies portability and interoperability risks. Atlas proposes a switch-mid-task test but has not conducted it.

Model access does not establish practical agent access

The ability to act depends on relevant language, tools and service infrastructure, not simply access to a foundation model.

Agentic Inequality is a conceptual framework; Nilekani supplies an attributed implementation perspective. No global prevalence estimate.

Reliability and user control are separate dimensions

Agents may become more reliable without becoming easier to switch; lower search costs can broaden choice if useful alternatives remain reachable.

Atlas’s scenarios and 2029/2031/2036 checkpoints are conditional, without probabilities or claims of inevitability.

A booking crosses several boundaries

Imagine arranging a hotel for a work trip. Today you might compare tabs, read cancellation rules and complete checkout. An agent could take over that coordination: retrieve information, narrow the options, request your approval and call a booking service. The hotel still supplies the room; someone must still confirm and manage the booking.

The illustration separates reading, choosing and acting. A page being readable does not make its rooms bookable through your agent. Permission to make one purchase does not authorize a different purchase. And a fluent “cancelled” message is not a cancellation record.

The internet does not disappear in this picture. Its visible interface becomes smaller while the relationships behind it become more consequential.

Power starts before the recommendation

Hold the traveller’s needs and the agent’s reasoning constant. Now remove a booking connection. A room can disappear from the usable shortlist without becoming worse, more expensive or less relevant. The system may still choose the best option it can verify. The restriction sits one step earlier: which options count as available.

That is why a better model alone cannot open this market. Services decide what information and actions to expose, on what terms and to whom. Authentication, paid access, commercial agreements and technical integration can shape the reachable market before ranking begins.

There is a real counterforce. In The Coasean Singularity?, Peyman Shahidi and coauthors argue that cheaper search and coordination could make comparisons worth doing that people currently skip. They also distinguish agents that travel across platforms from agents supplied by a platform: deeper integration can improve convenience while limiting outside options. This is economic analysis, not a measured welfare gain; coauthor Andrey Fradkin works at Amazon.

For the traveller, a tightly integrated agent may be the sensible choice when it has the right room and handles changes reliably. Openness matters most when an outside alternative is useful and switching to it actually works.

A helpful agent still needs a business model

The traveller wants a suitable stay with little effort. A hotel wants a booking at an acceptable margin. An intermediary may want a commission, a subscription or a continuing customer relationship. Those interests can align. They can also diverge when the best outcome for the traveller is to choose another supplier—or buy nothing.

Nithin Kamath, who runs Indian brokerage Zerodha, makes this conflict unusually concrete. In a company-published interview, he welcomes customers building their own interfaces over broker APIs, but expects other brokers to resist because they would lose opportunities to prompt customers to trade. That is his operating perspective, not a finding about every broker. It explains why technically possible access can be commercially unwelcome.

The funding routes below are possibilities, not a survey of current agent products. Ask what is being rewarded and what the user can inspect. A subscription is no guarantee of loyalty; a commission is no proof of a bad recommendation.

The agent needs information someone still has to produce

Our traveller may need a local review to learn that a hotel’s “quiet room” faces a late-night venue. If the agent supplies that answer without a visit, the review can remain useful while its publisher loses a chance to earn from the reader. That is a possible economic tension, not a claim that every answer replaces a visit.

Pew observed traditional-result clicks on 8% of Google visits with an AI summary, versus 15% without, among 900 US adults in March 2025. The study concerns search behaviour; it does not establish what an autonomous travel agent will do.

One possible response is to charge for machine access. Cloudflare’s open agent internet proposal includes paid requests and publisher control over access. Cloudflare sells infrastructure that could support that arrangement, so its proposal is a commercial direction rather than independent evidence of a sustainable publishing model.

Atlas’s inference is a feedback problem: if agents reduce valuable visits, some sources may restrict access or seek payment; those terms then affect what agents can consult. A workable replacement must fund useful information without making only the easiest-to-license sources visible. More referrals, subscriptions or new paying demand could weaken this pressure.

Control matters most after you click yes

The booking is approved. Who may charge the traveller, who confirms the order, and who can reverse it? These are different roles. A recommendation is not a payment mandate, and a payment token is not proof that the requested room was booked.

The August 2026 Universal Commerce Protocol specification separates the platform, merchant and payment credential provider. In its payment design, the platform passes a token to the business, which completes the transaction; the platform need not handle raw payment credentials. An optional extension provides a way to carry verifiable authorization. This documents an architecture, not its adoption by hotels or reliable performance in this example.

Now the meeting is cancelled. Useful control means retaining the terms and confirmation, being able to revoke further authority, and having a route to cancellation or dispute. Revoking an agent’s access does not itself undo a booking. Nor does moving a chat history give a new agent permission to manage it.

The UK Competition and Markets Authority identifies data mobility and interoperability as conditions for switching rather than becoming locked into an agent ecosystem. Its analysis also warns about commercially motivated steering. These are identified risks, not measured outcomes across global agent markets.

The practical test is whether the traveller can change agents halfway through the job without losing the evidence or ability to finish it. That remains a test to run, not a capability Atlas assumes.

The same chatbot does not mean the same power to act

A service available in two countries may support local payments, identities and merchants in only one. A traveller can receive an excellent explanation in their language yet still have to do every consequential step manually. Comparing model access alone misses that difference.

Matthew Sharp, Omer Bilgin, Iason Gabriel and Lewis Hammond’s Agentic Inequality framework separates availability, quality and quantity of agent access. Its September 2026 manuscript argues that access to a foundation model does not ensure access to a functioning agent. This is a conceptual framework, not a global measurement; the authors span Oxford, Google DeepMind and the Cooperative AI Foundation.

Nandan Nilekani offers a complementary implementation priority: useful applications and local-language access rather than fixation on the largest model. The Infosys cofounder also supports AI4Bharat. His position helps sharpen the question for our trip: can the agent use the local service the traveller needs, or only describe what they should do?

Across the US, Europe, India and other markets, the useful comparison is the completed route: language → relevant information → authorized payment → a way to resolve problems. Progress in one link does not establish progress in the others.

Two questions separate the possible futures

Separate two questions: can agents complete consequential work reliably, and can users move between providers with useful context and access? Neither implies the other. The same world can contain all four arrangements below, in different sectors and countries.

By 2029—three years from this article—watch ordinary exceptions: a price changes, a connection fails, a booking must be cancelled. Do agents complete the job or hand it back? By 2031, watch switching: can another provider take over a live task without rebuilding everything? By 2036, watch bargaining power: can suppliers reach customers through several viable routes, and can people refuse an intermediary’s terms?

These are checkpoints for conditional scenarios, not deadlines or probabilities. Evidence of reliable cross-provider completion and usable exit would support the more open path. Better task performance combined with exclusive access and costly switching would support the concentrated path.

For a business, the strategic question is which relationships remain portable: customer facts, service access, payment and fulfilment. Improving discoverability helps only if the route from consideration to completion is usable. The companion article, “Why would an AI agent choose your company?”, examines that selection problem in detail.

For the traveller, the promise is less work. The test of progress is whether that convenience comes with a real choice about who acts, what they can do and how to leave.