The agent runtime is the new layer of the browser, and your website will be evaluated against the runtime, not against an individual model.
This is a change that web professionals have not yet made. The conversation always revolves around models. Which model writes better? Which one cites the most precisely? Which API is the cheapest this month? The model conversation is loud as new models ship every few weeks and every release hits theaters.
The interesting story is the one below. The foundation is being rebuilt. This week it was impossible to ignore.
The runtime stack shipped in April
On April 15, Cloudflare delivered the Think projecta new Agents SDK built around durable execution with crash recovery and checkpointing, sub-agents that run as isolated children, persistent sessions with tree-structured messages, and sandboxed code execution running on Dynamic Workers. In the hours following the same day, OpenAI has delivered the next evolution of its Agents SDK with a native sandbox execution and a native harness of the model. Two of the largest infrastructure operators on the web have provided competing answers to the same question, with the question being: How does a long-lived AI agent actually perform in production?
Then, on April 16, Cloudflare added five more items. AI Platform: a vendor-agnostic inference layer that routes models to agents. AI research: a vector index and segmentation pipeline delivered as a managed product specifically for agent recovery, competing with Pinecone and Algolia in the agent-side RAG layer rather than Google AI mode. Email service in public beta, designed for agents to use the world’s most universal interface as a channel. PlanetScale Postgres and MySQL in Workers. And the technical basis for hosting very large open source LLMs like Kimi K2.5 directly on the Cloudflare network.
Sundar Pichai described the same change a week earlier. On the Cheeky Pint Podcast April 7 along with John Collison, co-founder of Stripe, he called Search itself an “agent manager”: “A lot of what are just information search queries will be agent in Search. You will complete tasks. You will have many threads running. The number of threads per query is a description of search execution. The Google CEO points to the same Cloudflare and OpenAI substrate delivered this week.
If OpenClaw was the web agent for consumers (a playable demo, an interesting prototype, enough to make a move), this is the web agent for adults. Sustainable. In a sandbox. Verifiable. The type of infrastructure you would actually run a business on.
The prevailing pattern in all of this is one thing: execution time. Not the model. Not the mainstream chat app. Not the main slide. Runtime is the layer where agents are launched, kept for hours and days, considering file system access, network access and memory. The runtime is the layer that decides if an agent’s session survives a crash, if its subagents can be reasoned with, if the execution of its code is contained.
The wrong question and the new one
Web professionals have spent the last 18 months asking the wrong question. The question was: which AI model should we optimize for? ChatGPT or Claude or Gemini or Perplexity. Which quotes matter most? Who should we let the robot pass to? This conversation made sense when models were reading your website directly.
This is no longer the case. The model reads what the runtime gives it. The runtime has retrieved your page. The runtime analyzed it. Execution time executed (or did not execute) your JavaScript. The runtime has resolved your structured data. Authentication negotiated at runtime. The moment the model sees something on your website, it sees the runtime’s interpretation of it.
The new question, if you take this week seriously, is what agent runtime environment your website is readable on. Three things to try before next week:
- Do your most important endpoints return machine-readable structured responses, or do they only display correctly in a full browser session?
- Is your authentication limited so that an agent acting on behalf of a user can hold a session across multiple calls, or does it only support one-time human connections?
- Does your structured data still mean the same thing if a runtime that didn’t execute your JavaScript tried to read it?
These are runtime readability issues. The model has nothing to do with them. The runtime decides whether your answer is even in the model’s popup, and the model selects what the runtime passes to it.
The web’s plumbing is being rebuilt. Every model for the next two years will see your website through one of these runtimes, not directly. Your website’s job, right now, is to be readable in runtime.
The model conversation will continue on the lecture stages and in the main slides. The runtime conversation happens in infrastructure companies’ product changelogs. The companies that ship the execution engine will decide which websites will be reached by AI search and commerce. Stop asking which model. Start asking which runtime.
More resources:
This article was originally published on No hacks.
Featured image: Viktoriia_M/Shutterstock
