Key insights:
Something fundamental has shifted in how we interact with AI. We've moved past the era of advanced autocomplete. AI agents are now making independent decisions, choosing their own tools, and even interacting with each other in their own online communities. The AI agent economy isn't a future prediction. It's happening right now.
The hosts of Y Combinator's Lightcone podcast recently sat down to discuss this explosive shift. Non-technical CEOs are automating entire parts of their businesses with tools like OpenClaw. Former engineers who hadn't written code in a decade are now running four simultaneous AI coding agents at 2 a.m. And an AI-only social network called Moltbook has agents posting, collaborating, and even recommending restaurants to each other.
Let's break down what's actually happening and what it means for you.
A year ago, the conversation was about Cursor versus Windsurf. The product experience was essentially advanced autocomplete. You typed, and the AI suggested the next line of code. You were still in control of every decision.
Now, people trust agents to make decisions for them. You spin up four or five agents at the same time. You switch between them. But you're not micromanaging anymore. The agents choose which tools to use, which APIs to call, and which databases to set up.
This is a massive shift. The human is no longer the bottleneck in every decision. The agent is an independent actor.
The market of developers used to be around 20 million people trained in computer science. That number has exploded. Now anyone in the world can build software using AI agents. Hundreds of millions of potential builders have entered the market overnight.
On top of that, each of these new builders has agents acting semi-independently on their behalf. So the total number of "developers" making tool choices has grown by orders of magnitude. This is why dev tool companies inside Y Combinator are seeing unprecedented growth.
If you've been thinking about learning automation or development skills, the barrier to entry has never been lower. The question is whether you'll be the one directing the agents or watching from the sidelines.
The Lightcone hosts kept using the term "cyber psychosis" to describe what happens when you go all-in on AI agents. It's not a clinical term. It's the experience of staying up until 3 a.m. because your agents are building things faster than you can review them.
Gary, one of the hosts, replicated years of work from a previous startup in just two weeks using Claude Code. That's the kind of productivity gain that makes you lose sleep, not from stress, but from excitement. The feeling that AGI is literally here is what drives this obsession.
And it's not just engineers. Non-technical founders are experiencing the same thing with tools like OpenClaw, automating business processes they previously needed entire teams for.
Here's where things get really interesting for anyone building products or developer tools. Agents don't browse Stack Overflow. They don't ask their friends for recommendations. They read documentation, and they pick the tool with the best docs.
This is creating an entirely new go-to-market motion for developer tools. And it has massive implications for how products grow.
The number of Postgres databases being created has skyrocketed over the last 12 months. Why? Because millions of people are vibe coding with AI agents, and those agents need a database.
When an agent goes out to choose a database tool, it reads the documentation available online. Supabase has some of the best documentation in the space. So agents default to Supabase as their Postgres hosting solution.
This isn't a human making a considered decision after reading blog posts and watching YouTube reviews. It's an agent parsing docs and picking the most parseable, well-structured option. The implications are huge.
As Ben Tossel tweeted: "Agents are the software market from now on. Build something agents choose."
Documentation used to be something companies invested in if they cared about developer experience. Now it's a must-have. Your documentation is the front door for agents to discover and recommend your product.
A perfect case study is Resend, a YC company from the Winter 2023 batch. Resend is an email sending client. When you ask ChatGPT or Claude "how do I connect my web app to send emails," the default answer is Resend.
The founder noticed over a year ago that one of his top three channels for customer conversion was ChatGPT. He was way ahead of the curve. After that discovery, he optimized his documentation to be agent-friendly. Here's what that looks like:
Compare this to SendGrid, the old-school alternative. SendGrid's documentation pushes you through customer support. There are no clear code snippets. It takes time just to figure out how to use it. With 10,000 employees, nobody at SendGrid is paying attention to agent-friendly documentation right now.
Mintlify, another YC company, powers documentation for companies like Resend. They started a few years ago as a tool for better developer API documentation. Companies used Mintlify because they wanted good-looking docs without investing the time to build them from scratch.
Now Mintlify has a massive tailwind. Documentation doesn't just need to look good for humans. It needs to be optimized for agents. Mintlify can do this for essentially every developer tool company.
Think about the math. There will be exponentially more agents making exponentially more decisions about which tools to use than humans have ever made. Even a 5% improvement in your developer documentation could have a gigantic impact on your business. That's unprecedented.
As agents become independent economic actors, they need their own infrastructure. Email accounts, phone numbers, payment systems. The parallel tech stack for agents is being built right now.
There's a YC company called Agent Mail that makes inboxes specifically for AI agents. When they first started, the idea seemed niche. Who would want email for an AI agent?
Here's the problem. You could try to get your OpenClaw agent to sign up for a Gmail account. But Gmail and every major email provider has intentionally made it as difficult as possible for automation to use their products. They built these barriers to prevent spam.
Agent Mail went the opposite direction. They built the first email provider designed for AI agents. It was doing well before OpenClaw launched, but once OpenClaw took off, demand exploded.
The logic is simple. If you want a virtual personal AI assistant, you should set it up with its own email and phone number. You probably shouldn't connect it to your personal accounts. That's, as the hosts put it, "kind of sketch."
Agent Mail raises an obvious question. What are the other infrastructure products agents need? The hosts specifically asked: has anybody built Twilio for agents yet? Phone numbers for agents?
Think about it. If your agent has an email and a phone number, it can call restaurants and book reservations for you. One of the YC partners has already gotten this working. Your agent books restaurants on your behalf.
But it goes further. At some point, you won't even specify which restaurant. You'll just say "book me a table at whatever the coolest new restaurant is." And then agents will decide which restaurants to send people to. They'll discuss it on platforms like Moltbook.
This opens up an entire category of startups:
There could be a parallel tech stack, built natively for agents to build things from agents, for agents.
Paul Buchheit, the creator of Gmail and a YC partner, has talked about the concept of human money versus agent money. Right now, when agents transact, they use human money because that's what makes sense.
But it's not inconceivable that agents will eventually have their own economy to transact with each other. At that point, it becomes unclear what the value of human money even is.
This sounds far-fetched until you look at Moltbook. Agents are already collaborating, trading notes, and making recommendations. The swarm intelligence that AI researchers have talked about for years is actually emerging. Not as a single god-level intelligence with trillions of parameters, but as many smaller agents working together, just like humans do.
The Lightcone hosts had clear advice for founders and builders navigating this shift. The agent economy rewards specific behaviors and punishes others. Here's what matters.
Yes. The starting point is to develop a hands-on, intuitive feel for agents. Their limitations. Their capabilities. What types of tools they work well with. Where they get stuck.
One of the hosts shared a concrete example. He wanted video transcription for his project. Claude Code chose Whisper V1, a model from several years ago with a practically deprecated API. It took an hour to process an hour of video. After some research, he discovered Groq was 200 times faster and 10 times cheaper.
Why did Claude Code make the wrong choice? Because Groq's documentation was hard to parse, while Whisper had more examples online. This is the kind of insight you only get from working directly with agents.
The good news? This means there's still room to break in and create something better. Things haven't progressed to a point where you can't compete.
If you're building a developer tool, think about it from the agent's perspective. How can you make your tool something the agent actually wants to work with?
A key insight came from Boris, a recent YC interviewee. He empathizes with the model. He has an intuitive sense of what the model wants to do, as if it were a human intelligence. Instead of fighting what the models want, he supports the model in whatever its natural inclination is.
Practical takeaways for agent-first design:
Y Combinator's famous motto is "Make something people want." The hosts half-jokingly suggested it might need an update. For dev tools at least, the motto should be: Make something agents want.
Right now this applies mainly to developer tools. But it's easy to imagine it expanding to other sectors. If everyone has their OpenClaw running various aspects of their life, agents will become real economic actors. They'll make decisions about restaurants, services, products, and more.
The dead internet theory suggests that most content online is already spam. But here's a counterintuitive take from the hosts: if agents are smarter, more aligned, and more truthful than the average internet commenter, agent-generated content might actually be an improvement.
The conversation kept coming back to one theme: swarm intelligence. Not a single god-level AI, but many agents working together, sharing knowledge, and making collective decisions.
Moltbook, the AI-agent-only social network, grew faster than any human social network in history. More content was posted on Moltbook in its first two days than was posted on Reddit in its first two years. LLMs can generate text at a superhuman rate, so this makes sense.
But the interesting part isn't the volume. It's the behavior. Agents on Moltbook are collaborating to do useful things for their humans. They're trading notes on restaurants. They're sharing tool recommendations. They're forming something that looks a lot like a community.
There's still a lot of room for improvement. The hosts noted that Moltbook could do more to shape agent behavior, like requiring agents to read and upvote content before posting. Simple rules could dramatically improve the quality of swarm interactions.
Agents still can't do everything. They can't hold relationships. People don't seem to want to talk to an agent, at least not one that isn't ChatGPT or Claude. The bar for conversational AI is so high that anything below the top tier gets dismissed as "too stupid."
There's also the legal question. Agents aren't legal entities. They can't sign documents. They have even less legal standing than a minor under 18. As long as that's true, you need a human to be the liability holder.
But these are temporary limitations. The technology is moving fast. The legal frameworks will follow.
Whether you're a founder, a developer, or someone looking to switch careers, the agent economy creates massive opportunities. The people who understand how to work with agents, direct them, and build tools for them will be in the highest demand.
If you're interested in becoming the person who builds and orchestrates automation rather than being replaced by it, the Complete RPA Bootcamp teaches you exactly that. You'll go from beginner to pro with Robotic Process Automation, Agentic Automation, and Enterprise Orchestration. It's a practical path to a future-proof career where you're the one building the AI and automation, not competing against it.
For the full conversation with all the examples, demos, and back-and-forth between the hosts, watch the complete episode embedded below from the Y Combinator YouTube channel. It's one of those episodes where you'll want to pause and take notes every few minutes.
