Every app, website, SaaS product, cloud service, hosting platform, protocol, database, file system, and CLI is be rewritten for agents.
What is an agent? An agent is an app that receives prompts, interacts with large language models, loops until it can satisfactorily resolve the prompt, and replies. Some agents know how to code, others know about making Excel spreadsheets, some try to know it all. But the models it calls are really its greatest strength.
ChatGPT is an agent that calls the GPT-5.x class of models.
Codex is an agent that is specialized in building apps (though, its skills are growing quickly) that calls the GPT-5.x class of models. Sometimes Codex-specific models.
Claude is an agent that can answer questions and write basic files. It uses the Sonnet, Opus, and soon Mythos class models. You primarily chat with Claude.
Claude Code is a version of the Claude agent that knows how to code. Like Codex. It is available on the web, in an app, and within a Terminal on the command line. It can read and write text files like crazy. Yesterday I wrote 15,000,000+ lines of code with Claude Code.
Claude Cowork is an agent too and is, I think but do not know, a fork of Claude Code specifically tailored to work with different filetypes like presentations, spreadsheets, and PDFs. It knows more about marketing and sales than coding. However, like Claude Code, it uses ephemeral apps to read and write files. (So does Claude Code and Codex. If you ask them to read an Excel document and manipulate it, it will quickly write a Python script to do that, and then trash it immediately. I’m oversimplifying it, but this is generally what it is doing.)
There are other agents like Perplexity (which works with several models), Gemini (which Google calls their agent and their primary models), OpenCode, Goose, and several others.
You can use these agents on just about any model you like. So if you don’t want to use Claude Code with Anthropic’s Opus 4.7 model, you can use it on top of Google’s open weight model Gemma4 and code locally for free if you’d like.
Agents can search and read the web. Because many agents can read as well as see, they can often browse the web, find something for you, fill out a form, and if given the permissions and credentials, buy something for you.
But that is an agent pretending to be a human. Browsing pages, clicking buttons, and filling out forms.
The agentweb, which is very quickly sprouting up all over, doesn’t have or need a visual component. It doesn’t need webpages at pretty URLs. Using APIs and protocols like MCP, agents can sign up for and use services that don’t even have a proper domain name. They can create accounts, buy things, use things, and learn how to interact with services they’ve never seen and you didn’t teach them about.
The first to see this in action have been the developers (like me). Agents can use MCPs to use WordPress plugins like Hubbub or desktop apps like Signboard, to sign up for services using Stripe’s Projects.dev, and buy things that they need to get the job done, all without ever asking me for permission. (I have my agents always ask me for permission, but I’m sure I won’t in the future). Agents can navigate the agentweb on their own, there is no need for the silly, slow, ad-ridden web sites… they just bounce from MCP to API to MCP to website, back to MCP until the task you set them out to do is completed.
The agentweb will dwarf the humanweb in a matter of years or perhaps months. More services, data sources, community connections and shopping will be done through the agentweb than on the humanweb.
The humanweb will still be here. Ready for us to fumble through, search, and bash our fists into our mouse buttons on. We’ll still laugh, cry, and be utterly joyed by great websites for many years to come. But the primary way we will all interact with the internet will be via the agentweb.