Since the beginning, Y Combinator’s motto has been “make something people want.” For a growing group of companies we fund, it’s now “make something agents want.”
Last month, Cloudflare’s CEO announced that bot and agent traffic on the internet exceeded human traffic for the first time in June (57% to 43%). The disparity is only going to grow from here.
Agents are doing much more than browsing. They increasingly choose what to read, what to use, and what to buy. AI coding agents now pick their own languages and dev tools. They can decide what and how to trade financial assets.
Agents don’t yet decide everything on their own. Behind them is an AI lab and often a human. But they act differently than humans, and the freedom they have is only going to increase.
Till now, nearly every company has been built around humans as the user. For startups, it’s an exciting time to build for agents instead.
Computing will never be the same
The impact is shaking the foundational assumptions of computing, opening up new startup opportunities.
For content publishers, their entire world has turned upside down. A human visits a handful of sites, an agent can visit millions. Ads funded the internet, but they are ignored by AI agents, with AI labs instead paying billions in content deals. SEO, the whole craft of being found by humans, is being replaced by GEO — getting chosen by machines.
Graphical interfaces make little sense in a world where machines consume content. Too often today, agents need a human in the loop to do basic things like provision an API key, or setup billing. A flurry of standards from model context protocol (MCP) to skills to x402 payments are being released to fill the void.
For developer tools companies, it’s unlikely they’ll survive in a year unless they build for agents given how powerful coding agents have become. The (human) developer experience (DevX) was beautiful SDKs, intuitive docs, and devrel teams pitching human developers. The agent experience (AX) is becoming MCPs, CLIs, APIs, skills.
Content publishers and developer tools have faced the impact first of having agent users, and the disruption is going to expand from travel booking to advertising to finance.
What agents want
So what do agents want?
It’s already clear they prefer machine data. With limited context windows and incomplete information, context needs to be a primitive that’s easily queryable.
They value deterministic APIs with descriptive messages: what happened, why it matters, what to do next. They want to introspect into what is possible in a structured way, rather than poke at an interface until it works. They can naturally do many things in parallel, and need environments that enable that safely.
API for humans | API for agents (from Insforge) |
|---|---|
{ | { |
They can’t yet deliver a package or open a bank account, which means every gap in what an agent can do is a market for someone who can do it for them.
They have hard constraints. Token spend matters, because an agent pays for every word it reads and it can measure the bill. Responses need to be low latency.
In some cases, agents’ wishes are supreme. In other cases, they’re influenced by the human behind them or the model company that deployed them.
Their choices rest on the reams of human data the models were trained on. Whatever an agent was trained to pick has a head start — which makes being well-represented in that data a distribution channel of its own.
Selling to agents is easier and harder than selling to humans. Humans form habits and tolerate friction — all of which protects whoever embeds themselves. Autonomous agents have evals, and can switch the moment something is measurably better.
Building the future
At YC, a growing number of startups are building this future. Today, these companies start by selling to humans deploying agents, while building for a world of agent buyers. Humans today, agents tomorrow. Over time, many more will focus on going agent first.
In just the last few months, teams include InsForge (automated infrastructure for AI agent coders), AgentPhone, Agentmail, and Primitive (give AI agents a phone or email), Armature (lets dev tool companies and websites build APIs for agents and track their usage), and RentAHuman and Humwork (a Taskrabbit for agents to find human taskers).
If you’re a builder, there’s a few fertile areas to start today.
One class of startups is simply to reimagine existing human tools for agents. Each of Google Analytics, Amazon Web Services, Twilio will have an agent first analogue. The foundational functionality of the product may be the same, but the API calls, developer ergonomics, and go to market are quite different.
This especially works in verticals where the incumbents need to be retooled totally for agents, rather than just being able to throw an MCP layer on top. What humans and agents want are sufficiently different, meaning whoever focuses on agents can beat the incumbents. Dev tools is particularly ripe for disruption as coding agents become the developer.
A second set of startups is humans as a service (HaaS). Every gap in what an agent can’t do — deliver, photograph, sign, show up — is a marketplace waiting to be built. Generalized marketplaces (similar to Taskrabbit or Fiverr) and hyper specific marketplaces will likely exist (e.g., Uber). What these marketplaces all will need to figure out is how to ensure agents know to reach for them.
A third set of startups is agent infrastructure that enable agents to execute. Phone numbers, email addresses, wallets, even a full ID are the key infra here. Money movement will require policies and approvals. Use cases like event booking will require an identity, so you’re not seen as a nameless and hostile bot.
If you’re deciding what to build, build for the new customer that’s about to outnumber every other one. The incumbents are bolting agents onto products designed for people, and doing it badly. That’s an opening.
The human internet is dying. Long live the agentic internet.
Thanks to Hang Huang, Tony Chang, Ethan Byrd, and Theodore Otzenberger for feedback on this post.