Back in July, on Meta's earnings call, Mark Zuckerberg said it was "extremely unlikely" that five years from now there won't be billions of people with a personal AI agent. Billions of agents, each working around the clock on behalf of one person. Follow that through and it is hard to avoid the conclusion: within five years there may be more AI agents browsing the internet than humans.
Think about that for a second. Just over 6 billion people use the internet today, roughly three in four people on Earth, according to Our World in Data. Every one of them could soon be sending software out onto the web on their behalf.
This offers a new way to look at how we build websites and communicate. AI agents will need to do the same things that humans do every day, and do them accurately, to make sure their outputs are correct and legitimate.
Machine to machine
Today, distribution and sharing on the internet are human to human. That may increasingly become machine to machine. This has never happened before. Purpose-built programs with set responsibilities will hand their outputs to other AI agents, which will then process and distribute the output of the first agent.
Search Engine Optimisation (SEO) is a huge part of how information gets found online right now. AI agents will play by different rules. They will likely reward accuracy, speed, and the age of information.
A quick check on the numbers
Zuckerberg's prediction was about people having personal agents, not about agents outnumbering humans. That second step is our inference. Billions of agents that never sleep, against 6 billion humans who do, is not a hard sum.
So what does this mean for websites today?
Well, we think this is why the llms.txt file is so enticing, so important, and potentially so valuable. It lets agents and humans understand the contract between a website and its audience, human or machine, and get information and data at source.
An AI agent will read the llms.txt to quickly understand a website's purpose and resources. It is a small file, in markdown, that website owners and administrators can maintain themselves to publish and share their most relevant pieces.
This is an opportunity to truly control how AI views the impact and usefulness of your website. We think that is super important for the future of the agentic web. We are already counting how much of the web has one.
It also means that computationally expensive methods, such as scraping the whole website to understand it, can be done periodically, if at all.
Put together, the fan-out of inter-agent requests and communication will be reduced, delighting the user with efficient, cost-effective outputs.
We believe in that future
That is why we are aiming to be the best llms.txt generator there is. Build yours today on llmstxt.studio and host it, then run the audit to see what agents can read.
Questions
Did Zuckerberg say AI agents will outnumber humans online?
No. On Meta's July 2026 earnings call he said it was extremely unlikely there won't be billions of people with a personal agent within five years. That agents could outnumber humans online is our inference from that.
What is the agentic web?
A web where software agents acting for people fetch, read and act on websites directly, so much of the traffic is machine to machine rather than human to human.
How does llms.txt help AI agents?
It gives an agent a small Markdown file at a known address that says what the site is and links its most useful pages, so the agent can understand the site without crawling all of it.