The agentic web is a proposed shift from browsing websites to delegating tasks to software agents. Instead of searching for a flight, comparing pages, filling forms and paying manually, a user could ask an agent to find a refundable flight under a set budget, confirm the important details and complete the purchase with permission.
That is the future Microsoft CTO Kevin Scott described in a 2025 interview discussed around Microsoft Build. It is an important technology vision, but not proof that the web has already changed. The practical question is whether websites can expose reliable data and actions that agents can discover, understand and use safely.
What Kevin Scott means by the “agentic web”
The conventional web assumes that a person is the operator. You open a browser, search or navigate to a site, read the information, make decisions, complete forms and click the final button.
An agentic web would move more of that work to software. An agent would interpret a goal stated in ordinary language, discover relevant services, query live information, compare options, ask for clarification and—where authorized—take an action on the user’s behalf.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →For example, the request “Find a refundable flight arriving before noon and book it if the total stays below $500” contains several separate jobs:
- Find airlines or travel services that can answer the request.
- Check current schedules, prices and refund conditions.
- Apply the arrival-time and budget constraints.
- Show the user the relevant options and total cost.
- Authenticate the user and obtain approval.
- Complete the purchase and return a receipt and cancellation information.
A conventional search engine mainly returns links. An agentic system aims to complete the workflow. The difference is not a chat box on a website; it is the combination of live data access, tool use, permissions and controlled action.
Why Microsoft thinks the web needs another interface
Much of the web was built for visual human consumption. Important information may be inside JavaScript applications, forms or inconsistent page layouts. A crawler can index a page, but an agent trying to extract a current price, account-specific balance or cancellation rule may encounter stale content, bot defenses or a design change that breaks its workflow.
Search engines also depend on crawling, indexing, ranking and centralized discovery. That model remains useful, especially for open-ended research. But it is not always the best way to obtain fast-changing or personalized information such as inventory, prices, appointment availability, delivery estimates or account data.
Scott’s argument, as summarized in contemporaneous coverage, is that websites could expose their own natural-language and structured interfaces. An agent could then ask a source directly for current information instead of relying exclusively on a giant index or scraping a page intended for humans. The relevant interview was associated with The Vergecast’s May 23, 2025 episode and the Microsoft Build 2025 period.
This is a vision and architectural direction, not evidence that the entire web has become agentic.
The three levels of an agentic website
Many discussions use “AI search,” “agents” and “automation” as if they were the same thing. They are not. A useful way to evaluate a proposed agentic experience is to separate three levels:
| Level | What the system does | Example |
|---|---|---|
| Answering | Finds and explains information from a site. | “Which checking accounts have no monthly maintenance fee?” |
| Assisting | Compares options or prepares an action. | “Compare these accounts using my expected monthly balance and direct-deposit needs.” |
| Acting | Executes an authorized operation. | “Open the selected account after showing me the disclosures and final terms.” |
A natural-language answer about waterproof jackets is not the same as placing an order. Likewise, an agent that explains a bank’s published fee schedule is not automatically capable of opening an account, moving money or changing a beneficiary.
What is Microsoft’s NLWeb?
NLWeb is Microsoft’s proposed and open-source approach for adding natural-language access to a website’s existing content. At a high level, it is intended to help a site provide AI-powered search and make its content available in forms that agents can consume.
That makes NLWeb better understood as a site-level natural-language interface layer. It could complement conventional navigation, search, structured data, feeds and APIs. It is not a universal replacement for search engines, and the dossier does not establish large-scale production adoption or current release commitments.
For a publisher, retailer or financial-information site, the attraction is direct access to its own content. A user could ask a site about account fees, mortgage terminology or budgeting resources without first translating the question into keywords and clicking through several pages.
The unresolved issue is whether such an interface gives the publisher more control or simply makes its content easier for large AI platforms to consume. A site may provide the authoritative information while another platform owns the user relationship, controls the interface and captures the commercial value.
Microsoft’s NLWeb project page is the appropriate place to check implementation details and current status rather than relying on commentary about the 2025 interview.
Where MCP fits—and where it does not
The Model Context Protocol, or MCP, addresses a related interoperability problem: how AI applications connect to external tools, data and services through a more standardized pattern instead of requiring a separate custom integration for every model and service.
That could help an agent connect to a budgeting database, a document system, a calendar, a commerce service or another tool. But MCP alone does not create an open agentic web. A protocol can standardize a connection; it does not guarantee that the connected service is trustworthy, secure, discoverable or economically viable.
A serious agent ecosystem also needs:
- Identity and authentication.
- Narrow, revocable authorization scopes.
- Tool discovery and reputation signals.
- Data provenance and freshness indicators.
- Rate limits and abuse prevention.
- Payment and settlement systems.
- Audit logs and dispute processes.
- Human confirmation for high-impact actions.
- Privacy and data-retention rules.
It is therefore more accurate to say that MCP can help an agent connect to a service. It does not automatically make that service safe or accountable. MCP was not created by Microsoft; its relevance here is its role in the broader conversation about agent interoperability.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
The official protocol documentation is available at modelcontextprotocol.io.
How the agentic web could change search and SEO
Search may gradually shift from “find pages” toward “help me complete a task.” For some queries, an agent could query several authoritative sources, compare their answers and return a result with citations, constraints and next steps.
That would increase the importance of:
- Clear semantics and consistent terminology.
- Fresh, structured data.
- Stable schemas and machine-readable constraints.
- Explicit descriptions of available capabilities.
- Reliable answers and source provenance.
- Well-defined authentication and permission boundaries.
This does not mean traditional SEO is dead. Search engines remain valuable for discovering unfamiliar sites, comparing sources, filtering spam and handling questions that no individual site has anticipated. Agents themselves also need directories, indexes, reputation systems or search layers to find services they have never encountered.
A more defensible interpretation is that agent readiness adds an API-design and trust-engineering dimension to SEO. The old question was, “Can a crawler understand and rank this page?” The additional questions are, “Can an agent understand this data, verify its freshness, identify its limits and safely invoke the relevant action?” Industry commentary sometimes calls this “agent compatibility as the new SEO,” but that is an interpretation, not an established standard.
Free tools Windows power users keep installed
One-click scans. No signup required.
What this means for publishers and financial websites
For publishers, banks, comparison sites and personal-finance services, direct agent access could have genuine benefits:
- Users could ask questions about complex products in ordinary language.
- Current rates, fees, eligibility rules and terms could be presented with more context.
- Large catalogs and databases could become easier to search.
- Businesses could expose controlled actions such as preparing an application or scheduling an appointment.
- Sites could make their own data available without relying entirely on search snippets.
But the risks are equally significant. If an agent summarizes an article or comparison without sending a visitor to the publisher, the publisher may lose referral traffic. The site still pays to create and maintain authoritative content, while a platform may capture the user relationship and monetization.
Financial information makes the problem sharper. An agent should distinguish a published educational explanation from a personalized recommendation, and a displayed interest rate from a rate a particular applicant is actually eligible to receive. It should surface fees, conditions, introductory periods, penalties and disclosures rather than compressing them into a confident but incomplete answer.
Publishers and businesses also need control over what agents may read, quote, cache or do. An “AI interface” without clear licensing, rate limits, opt-out controls and capability boundaries can increase costs and exposure without producing a corresponding business benefit.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchCan agents safely handle money?
Reading information is comparatively low risk. Taking action is different. A financial agent that moves money, applies for credit, changes account details or accepts a contract needs materially stronger controls than one that summarizes an article.
A safe transaction workflow must handle:
- Discovery of a legitimate service.
- Understanding of its capabilities and restrictions.
- Retrieval of current information.
- Application of the user’s exact constraints.
- Presentation of the options, total cost and material terms.
- Explicit approval for the consequential step.
- Strong authentication.
- Payment or account authorization.
- Confirmation, receipt and an audit trail.
- Cancellation, recovery and dispute handling.
Failure can occur at every stage. Inventory or rates may change between quotation and execution. A tool description may omit an important restriction. Authentication may expire halfway through a workflow. A retry after a network failure could make a duplicate purchase or payment unless the operation is idempotent.
Agents also face prompt injection and malicious-tool risks. Content returned from a website can contain instructions aimed at the agent rather than the user. A compromised or misleading tool could request excessive permissions. The agent must keep user intent separate from untrusted instructions in retrieved content.
For high-impact activity, the most sensible design is graduated permission: read-only access first, preparation next, and explicit confirmation before an irreversible action. The user should be able to see what will happen, which service will do it, how much it will cost and who is responsible if it goes wrong.
The “capability overhang” problem
The discussion around Scott also used the idea of a capability overhang: models may be able to reason, summarize, code and use tools faster than products and infrastructure can turn those abilities into dependable services.
That is a useful framing, but not a measured industry statistic. The bottleneck may be less about raw model intelligence and more about fragmented applications, weak permissions, unreliable integrations, unclear accountability and poor recovery procedures.
An agent can be impressive in a demonstration and still be unsuitable for production if it cannot explain the source of an answer, detect stale data, handle partial completion or obtain narrowly scoped authorization.
What exists now versus what remains speculative
| Layer | Already familiar | Still unresolved |
|---|---|---|
| Content retrieval | Search, feeds, APIs and retrieval-augmented systems. | Reliable cross-site meaning, freshness and authority. |
| Natural-language access | Chat interfaces and site search. | Consistent quality, citations and clear boundaries. |
| Tool use | Function calling and connectors. | Universal discovery, permissioning and trust. |
| Transactions | E-commerce and financial APIs. | Liability, payment, identity and recovery across services. |
| Autonomy | Narrow, supervised workflows. | General-purpose delegation that users can safely trust. |
The demonstrated pieces are real and increasingly familiar. The speculative leap is a broad, open web where agents routinely find unfamiliar services, negotiate capabilities, authenticate, transact and recover from errors without a person navigating every step.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesFour questions for judging an agentic-web proposal
- Can it find the right source? A decentralized interface still needs discovery and reputation.
- Can it prove the answer is current and authoritative? A fluent response is not evidence of accuracy.
- Can the user control the action? Permissions should be narrow, visible and revocable.
- Can someone recover from a mistake? Receipts, logs, cancellation paths and liability matter as much as convenience.
These questions also expose the main trade-offs. Centralized search offers broad discovery and spam filtering, while site-level interfaces may offer fresher and more domain-specific information. Scraping works with existing sites but is brittle; APIs and tools are more reliable but require investment. Greater autonomy saves time but increases the cost of errors. Open protocols may reduce dependence on one provider, yet major platforms can still control distribution, identity, billing and model access.
What Scott’s vision gets right—and what it does not prove
Scott is describing a plausible evolution of the web’s interface layer. Websites are likely to expose more structured data, natural-language search and machine-usable capabilities as models become better at tool use.
That does not prove that browsers, search engines or websites will disappear. Search remains important for discovery. Browsers remain useful when a user wants to inspect a source directly. Websites remain the place where businesses publish information, establish identity and enforce their own rules.
The harder question is economic and institutional: who owns the relationship when an agent becomes the interface? A site may provide the data, a model provider may perform the reasoning, an agent platform may control the workflow, and a payment or identity provider may authorize the result. “Open” at the protocol layer does not necessarily mean open competition or stronger publisher control.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →For users, the best version of the agentic web would make routine work easier while preserving visibility into sources, prices, permissions and consequences. For businesses, it would create useful machine interfaces without turning every customer interaction into an unaccountable black box. That is the standard the technology will have to meet.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
