How I see the agentic web evolving

AI is moving faster than any of us can reasonably keep up with. While everyone is busy adapting to the changes we’re already seeing, it feels nearly impossible for most teams to imagine what’s coming next. And because the pace of change is so fast, what’s coming next is closer than you might expect.

I’m already on record stating that every website, every app, every digital experience now has to be two things at once: human-first and AI-native.

By “human-first,” I simply mean that no matter what technology sits in the middle, the beginning and the end of every digital journey still have to serve great human experiences.

• At the beginning, you still have a creator: People with a vision, a story, or a product to share. 
• At the end, you still have a consumer: People looking for an answer, a product, a service, or a solution.

AI-native is what’s constantly changing that middle space. We’ve already seen that interfaces are becoming more conversational, content is becoming machine-readable, and platforms are becoming machine-operable. This is where all the opportunity (and all the risk) live. For businesses, the problem will be finding an AI-ready foundation that doesn’t require you to scrap the site you’ve already invested years into building.

With that context in place, I have a few predictions about how agents will interact with websites in the very near future. If I’m right, you’re going to feel the weight of these changes on your infrastructure.

The emerging stages of the agentic web

Most of the conversation right now treats AI citations in LLMs and AI Overviews as the finish line for “AI readiness.” I think it’s just the start of the race, and we’re about to see two more stages emerge. Here’s how I see the agentic web evolving:

1. Training data stage: This is where we’ve been operating for the past few months. Your content is used to train the model, and the model recalls your answers when someone asks a relevant question. 
2. Live agentic search stage: This has already started and is growing quickly. Instead of recalling what a model absorbed months ago, an agent searches the live web on a user’s behalf and reasons over what it finds. Parallel, founded by former X CEO Parag Agrawal, is a good example of this. They’re building web search infrastructure designed specifically for agents, with results already structured for machine reasoning.
3. Just-in-time execution: This is what I see on the horizon, and it seems all but inevitable. Sites are starting to expose actions an agent can take, not just content it can read. It’s what will let someone research, decide, and actually purchase an item or book a trip without ever leaving their ChatGPT or Claude interface, and do so safely, from a source they trust.

Making predictions about future agentic standards

New agentic standards are starting to emerge and are at various stages of adoption. For example, two standards that will be crucial for that third stage, and that are being widely discussed and increasingly adopted, illustrate this well. 

The first is WebMCP, an open standard proposed mainly by Google and Microsoft engineers that lets web pages expose callable tools an agent can invoke directly. Actions might include searching a catalog, booking a consultation, adding an item to a cart, etc. In late August, OpenAI added WebMCP support to the ChatGPT desktop browser, so a visiting agent can discover and use whatever tools a site offers. This is moving fast, but the spec is still a draft community group report, so it’s not a W3C standard yet.

The second is UCP. Created mainly by Google and Shopify, it’s how merchants can grant agents the ability to discover products and complete checkout on the spot on behalf of their users.

Don’t be fooled by low adoption rates now. Think about it this way: WebMCP has about a 4% adoption rate right now. That number doesn’t seem significant, but before OpenAI announced native support for WebMCP back in August, the adoption rate was only at about .001%. Moving from .001% adoption to 4% is a huge leap, and it happened in about a month.

AI readiness will involve several new standards and requirements. For some, we can clearly see active adoption, but for others, we’ll have to rely on informed perspectives on future growth.

Where I think this is heading

The interfaces mediating all of this (ChatGPT, Claude, Gemini, etc.) want to keep users inside their own surface. They’re deeply invested in building a web where agents can find the information and do the tasks a user requires without ever sending them elsewhere. To me, it seems natural that these interfaces will start rewarding sites pre-structured for agent operability.

Search engines have already done something similar: sites that provide better experiences for people tend to rank higher. My bet is that AI interfaces will do the same thing for agentic experiences, favoring sites that are machine-readable and operable.

I’d also bet that an agent’s preference for sites where it can conduct live searches and take direct actions will outweigh those optimized for training-data scraping sooner than most people expect. Relying on a snapshot of the web that goes stale the moment it’s captured only makes sense because live agentic search capabilities are still maturing. Once those standards have been implemented at scale, why would a model reason from an old copy of your pricing page rather than checking your current one directly?

If you’re planning your website strategy for the next two years, this progression from citation only to live agentic search to agentic execution is the future I’d prepare for.

What this future costs

This agent-focused future comes with a real bill attached. We all know agents are consuming more of your site resources, and I would guess that each of the three stages I outlined will become more expensive than the last.

A standard crawler grabbing your content for training is a relatively brief, cheap transaction. An agent searching the live web and reasoning over your pages asks quite a bit more. An agent executing a task will need the ability to keep a session open, call tools, wait for responses, and return for more as needed. That’s going to take much more from your website in terms of resources.

We’re already seeing the request volume expand, and we’re still at the beginning of phase two. Cloudflare Radar found that bots already generate more than 60% of HTML requests as of this September, and most of that is still just crawling. Imagine how much higher that will be when agents are regularly conducting sustained, persistent work against your origin servers.

AI isn’t causing website problems; it’s encountering infrastructure ones

Open source platforms like WordPress®, Drupal, and Joomla are built on PHP and MySQL. It’s an excellent stack, and I’ve spent years of my career working with it, but the hardware running it was never designed for these long-running, persistent operations. 

We’ve solved a version of this problem before for some specific needs. Years ago, we stopped storing large images and videos on web servers and handed resizing and delivery off to infrastructure built specifically for it. Nobody called that an open-source failure. We just needed specialized infrastructure to keep up. The same logic applies here: asking the machine that serves your homepage to also carry heavy AI workloads slows down both jobs at once.

This shows up most clearly when you see it at scale. One agent making a simple request to a typical PHP site works fine today. When thousands of agents start making sustained, reasoning-heavy requests around the clock, the cracks in the stack start to show.

Solving the infrastructure gap

The real fix is to once again move the heavy work off your origin server and onto infrastructure built for it.

Semantic search showed how this works, and RAG (retrieval-augmented generation) extended it. Vectors let AI connect related concepts and understand site context, and RAG uses that retrieval to ground AI-generated answers in a site’s own content. But a typical WordPress origin and its MySQL database aren’t built to store and run similarity search over large volumes of high-dimensional vectors without hurting site performance, so purpose-built infrastructure emerged to host, scale, and secure that vector data separately.

As we move toward stage three, I see a few critical requirements for specialized AI infrastructure that lives alongside your site: 

1. Vector storage and retrieval, so AI can understand your content.
2. LLM inference available directly within your CMS, so you can build with AI that’s grounded in your site’s specific context.
3. Cloud-hosted MCP, so AI can take action on your site without slowing your website’s origin server. 

That last piece matters most for what’s coming: MCP gives an agent the legs to come fetch things from your site, and those long-running sessions need a foundation that won’t buckle under the weight. 

With this setup, your business-critical sites continue to run on the platform you’ve already invested in; you just add the capacity needed to serve agents.

Start placing your bets

If you consider your website a critical business asset, it’s time to bet on agentic search and just-in-time execution as inevitable. Start looking into the infrastructure that can support them right now, before heavier traffic and higher expectations show up at the same time.

For me, this whole conversation perfectly encapsulates why I keep betting on the open web. There are so many individuals and organizations working to solve these problems together, so when a new model, protocol, or interface emerges, you don’t have to jump ship and rebuild from scratch. You just need to look for the infrastructure that’s already solving the problem, plug it in, and keep building.