Welcome to the new era of AI agents: the future is not "created for humans," but "serving AI."
```
When AI agents start “making decisions for people,” they will reshape software distribution and usher in the next generation of infrastructure.
Recently, on the latest episode of the podcast "Lightcone," Y Combinator Managing Director Jared Friedman, Managing Partner Harj Taggar, and General Partner Diana Hu discussed the changes brought by AI agent tools like Claude Code and OpenClaw. They believe that as “agent experience without human intervention” matures, the entry point for developer tools is shifting from human search and community reputation to “what agents recommend.”
They mentioned cases such as Resend and Mintlify, suggesting that documentation/knowledge bases will become the key entry point for LLMs to recommend tools: “The optimization target is no longer humans, but agents.” At the same time, “agent-native” email, phone numbers, and even a future “agent economic ecosystem” are starting to emerge, as seen with tools like Agent Mail.
The “Agent Experience” Without Human Intervention Becomes a Watershed
According to YC, the core change is not “smarter autocompletion,” but users beginning to hand over decision-making power to agents. Harj Taggar said directly: “The key is no human intervention... people fully trust agents to make decisions for them.” He emphasized that when several agents run in parallel and users only need to switch between them, “you no longer need to manually manage them.”
Focusing on the competitive landscape of developer tools, YC highlights a key judgment: the “default option” for software is shifting from human developer communities to recommendations and invocation paths of LLMs/agents. Jared Friedman quoted a tweet: “From now on, AI agents are the software market.” Diana Hu then adapted YC’s classic slogan “Make something people want” to a more direct version: “Build what AI agents want.”
In their narrative, the reason tools are chosen is also changing. In the past, it was more like the outcome of “developers talking to each other”—“Stack Overflow... or popular code repositories created by people on GitHub”; but now, the developer market is rapidly expanding thanks to AI: “from about 20 million developers trained in computer science, expanding to anyone in the world can now be a developer.” The host added: “Now it might already be hundreds of millions.” On the other hand, agents are also becoming new “users” and “decision-makers.”
Documentation/Knowledge Bases Become the “New Placement”: Optimization Target Shifts From Humans to Agents
In YC’s discussion, a frequent conclusion is: documentation and knowledge bases are becoming the key entry point for agents to recommend tools.
The guests mentioned the case of the email tool Resend: When Resend launched its email sending client in Winter 2023, “if you ask ChatGPT or on the cloud, or on almost any mainstream platform, ‘how do I connect my web app to an email sending system,’ the default answer is actually Resend.” More importantly, the founder spotted the channel change early: “He posted that ChatGPT was the top three channel for customer conversions. Afterwards, he focused on optimizing documentation to make it more usable for customer service staff.”
They also specified standards for “agent-friendly documentation”: knowledge base content organized around “questions users or customer service might ask,” answers “clear and structured,” with plenty of copy-paste-ready code snippets—“it turns out this is perfect for LLM and robotic operation.” By contrast, the host criticized traditional tools: “It’s like you’re being sent straight to customer service. Where are the code snippets? I don’t even know how to use it.”
They also called out the logic for document tools like Mintlify: In the past, docs were “a bonus,” now they’re a “must-have,” because “docs need to be optimized, but the target is no longer humans, but agents.” Harj Taggar summed up the business meaning: “Even if you can barely improve developer docs by 5%, for your business... the impact could be huge.”
Agent-Native Infrastructure Emerges: Email, Numbers & “Agent Economic Ecosystem”
As agents take over more execution, YC believes there will be a set of “agent-native” infrastructure. Jared Friedman mentioned YC company Agent Mail: It “developed inboxes specifically for AI agents,” because traditional email systems “deliberately set up many barriers to make it hard for automatons to use their products, to prevent spam.” After OpenClaw rose, “things changed,” and Harj Taggar describes its growth as “explosive.”
Discussing “agent identity” and contact methods, Harj Taggar even offered “best practice” advice: “If you want a virtual personal AI assistant, the best way is to have it set up its own email and phone number.” Guests also raised bigger ideas: could there be an “AI-agent version of Twilio,” even a “tech stack entirely built for agents, enabling agents to build apps for agents.”
Extending to transactions and payments, Harj Taggar mentioned the possibility of “human currency and agent currency”: “Maybe, eventually, they will have their own economic system and trade with each other... at that point the value of human currency will be hard to determine.”
The following is a full translation of the interview:
Jared Friedman-- Y Combinator Managing Director
Harj Taggar-- Y Combinator Managing Partner
Diana Hu-- Y Combinator General Partner
Host 00:00
Welcome to another episode of Lightcone. Things have gotten a little different lately. First, Claude Code has completely taken over my life. If Jared’s situation is any reference, I suppose OpenClaw has taken over his life too.
Jared Friedman 00:15
I’m obsessed with a new site called Moltbook, where people unleash their AI to interact in the first fully agent-only online community. I’m even here simulating my personal OpenClaw instance.
Host 00:29
Okay, guys, I can’t. We have to bring it up.
Host 00:40
Okay, we’ve said it all. I mean, so many crazy things are happening. Some of my non-technical CEO friends are all in on OpenClaw right now. They’re fully automating parts of their business with OpenClaw, it’s insane. Meanwhile, there are product CEOs and former engineering CEOs—like me. I haven’t written code in ten years. Yet these days, I pull all-nighters running four parallel jobs with cloud code. The model capabilities are exploding. We’ve talked about this for years, but now it actually feels like AGI is really here. We’re on the edge of transformation. Almost everyone knows somebody deeply addicted to online worlds—and I’m one of them. So, guys, what is happening now? You said you’re fully devoted to model making. What’s going on?
Jared Friedman 01:41
Yeah, you’re totally real. Feel the AGI moment. Gary’s moment was like using cloud code to build a whole startup, copying years of startup work in two weeks, it’s unbelievable. I feel the same, that moment of AGI, reading those AI cases, watching them interact as if they live in their own world, hardly needing human intervention, made me realize: in a few years, when these agents are released to live independently without humans, what will the world be like?No Human Intervention Is Changing the Experience
Harj Taggar 02:12
I think the key is no human intervention. If you recall a year ago, we were talking about cursor and windsurfing, and the product experience was basically advanced autocomplete. Now, clearly, with “call code,” people fully trust agents to make decisions. Like you said, four or five different agents running simultaneously, you can switch between them, but you don’t have to manually manage them anymore. This means agents can autonomously choose, say, post their own content on sites like Moltbook, which is an interesting and unexpected application. More interestingly for developers, agents can choose the tools used for building, actually spawning a complete agent economy, where agents pick development tools, products, goods, and services. Who knows? You’re seeing a full agent economy parallel to human economy.
Speaker 4 03:09
I think, before all this, developer tools were chosen more as a result of developers talking to each other.
Host 03:21
Overflow, Stack Overflow, reputable.
Speaker 4 03:23
Right? Or like popular human-made code repositories on GitHub. But I think the landscape for developer tools is changing drastically for two reasons. First, like you mentioned, developer markets suddenly expanded from about 20 million people trained in CS, to anybody in the world can now be a developer.
Host 03:46
Now it could be hundreds of millions.
Speaker 2 03:48
Plus all their AI agents, acting independently, as I just said.
Speaker 4 03:53
And then there’s those agents acting as prophets, telling you what tool is best. We saw these trends in YC incubated companies and developer tool companies, thriving because of these trends. Maybe we should talk about these trends and their underlying reasons.
Harj Taggar 04:10
Right? Immediately a statue-like example comes to mind: Azerisk Love’s friend mentioned, if you look at the number of databases created in the past 12 months—simple databases, Postgres databases—it’s exploding because all VIP programmers are coding/building apps, and choosing Superbase as default for setting up and hosting Postgres. YC companies see a chain reaction. Database demand exploded. The interesting part is that programmers pick Superbase as the default tool because its documentation is best, so they assume it’s best.Should YC Change Its Motto?
Jared Friedman 04:54
Use. There’s a great tweet about this: maybe we can post it too, and point out that from now on, agents are the software market. Developing a product agents would choose, that leads to a possibly controversial topic—should we change why she became a model?
Diana Hu 05:10
Build what agents want.
Host 05:12
For developer tools, just like the first day on the job you get a special T-shirt.
Jared Friedman 05:17
Now it’s just developer tools, but I can imagine in the future it could become a major force in other economic fields. If everyone can control their own lives, these agents would become true economic participants. They’ll ultimately make a lot of decisions.
Host 05:30
Yeah, interestingly, I had a personal “early days” moment—since I was building Gary’s list. One feature I wanted was video transcription. Often, eg with con convention content, the only way for my language manager to understand is with a transcript. Transcripts are usually missing, so I have to download, send to Whisper etc. You know, Cloud Code chose it for me. But it picked Whisper v1, from years ago, API basically deprecated. I sat there troubleshooting: why does processing an hour video take so long? Shouldn’t it be faster? Turns out, it takes an hour. I was stumped. Afterwards, I realized I shouldn’t use this model, I should use Grok with a queue. It was 200x faster and 10x cheaper. So this is a neat example of cloud code not optimized yet. Otherwise, two weeks ago it wouldn’t have happened, which is good—cloud tech hasn’t reached an absolute wall, you can still make better stuff.
Speaker 4 07:01
But I think this example flags another issue, Gary. Part of the problem is Grok’s documentation is really hard to understand. So, yes.
Host 07:10
Through.Email Tools & Agent Infrastructure
Diana Hu 07:11
Compared to Whisper, Grok is more appropriate, with more use cases. This is changing developer tool marketing. A concrete case: Resend finished a batch of email sending client dev in Winter 2023. If you ask “how to connect my web app to email sending system” on ChatGPT, Cloud, or most mainstream LLM platforms, the default answer is Resend. The founder noticed ahead of the curve: over a year ago, he posted ChatGPT was one of top three channels for customer conversion. Then he focused on optimizing docs for customer service staff.
Host 08:01
Yeah, what does it look like?
Speaker 4 08:02
One of the best things lately is optimizing in many aspects. If you look here at the knowledge base, many usages are targeted to questions users or customer support might ask, like “how to send or receive emails?” Clicking, you get clear, structured answers.
Host 08:23
I had this issue today. I wanted my program to receive emails and run on the cloud. Searched the web—no answer. Went to Perplexity, typed “RES resend function to help me receive emails?” Copied and pasted the reply, and it worked.
Speaker 4 08:41
The best part is tons of examples. He really made lots. In the code, each example has code snippets, support staff can check one by one, and the structure is very clear. Turns out, this is perfect for LLM and bots. LLM’s textbox is optimized to help support staff promote default resend functionality. Compare to SendGrid and similar tools, SendGrid is classic Web 2.0, not as easy to use. It just sends you to customer service. Where’s the code snippet? I don’t even know how to use it. Just understanding takes time.
Host 09:30
Yeah. Probably 10,000 people work there. But now nobody seems to care. I...Agent-Driven Documentation
Speaker 4 09:36
This leads to another point—docs are the key to recommended developer tools by many agents. I think one company doing interesting developer docs is minify that you.
Speaker 3 09:49
Collaboration is tough. Hey, I think minify is the behind-the-scenes hero of recent docs, right? Yes, minify is an interesting case. Years ago, my initial goal was a more complete API dev tool doc, and there was a market need. Yes, dev tool companies basically all use minify because they want good-looking docs without spending too much time. Minify has basic practical features, like auto-extracting API updates to update docs. Minify developed rapidly. Always did well, but now it’s a tailwind—docs’ importance is shifting. Previously, companies might glance at docs, especially those focused on design/dev experience. Now, docs are critical to all because docs need optimizing, and the target is no longer humans but agents. So nearly all dev tool companies do this. If you push this trend, you see an exponential growth of agents needing to make more decisions about which tools to use than humans ever did.
Harj Taggar 10:56
You know, even if you can barely improve developer docs by 5%, for your business, dev tool impact could be massive, unprecedented.
Jared Friedman 11:08
Speaking of email, there’s a YC company closely related called Agent Mail. They developed inboxes specifically for AI agents. When they first entered, it seemed very cutting-edge, didn’t know who would need it. Makes sense: theoretically, you could have your OpenClaw bot register a Gmail account for email, but it’s really hard, since Gmail and all providers deliberately set up barriers so automatons can’t use their products, to prevent spam. So Agent Mail does the opposite. They built the first email service for AI agents, before OpenClaw even existed, and it was doing well. But after OpenClaw’s rise, things changed.
Harj Taggar 11:53
Explosive growth. That’s the perfect example, right? I know some people connect OpenClaw to their personal emails, but that’s not...
Host 12:01
Like a sketch.
Harj Taggar 12:02
Yeah, you shouldn’t tell anyone or mention it on Twitter, but if you want a virtual personal AI assistant, best to give it its own email and phone number.
Jared Friedman 12:14
Has anyone developed a Twilio-like dedicated AI agent service, or dedicated AI agent phone numbers? I know agent email makes me think: What other contacts do agents need besides email?
Diana Hu 12:23
Build? Sounds like pitching to a startup. Maybe there’s a parallel world with a tech stack that’s fully for four native agents, letting agents build apps for agents.
Harj Taggar 12:34
This maybe relates to Jared’s earlier point—not just dev tools. People often don’t want to book restaurants themselves. Now, if your agent has your email and phone, it can call for you. I think an IC partner Anchor is already doing this. Your agent can book a restaurant. You might trust it enough to say: “Hey, book me a seat at any new restaurant.” Then agent decides which restaurants to recommend on MoltBook. We’ve crossed some uncanny valley to get here.Collective Intelligence
Jared Friedman 13:28
Future leaders. 100%.
Harj Taggar 13:29
Takes me back to Paul Buchheit’s recent comment, always good at predicting. He mentioned the idea of human currency and agent currency, which might be where things are headed. Right now, agents make trades. If you let them trade, they’ll use human currency—it’s rational. Maybe someday they’ll have their own economy to trade among themselves—it’s possible. At that point, human currency’s value would be uncertain.
Host 13:58
YC’s next batch is accepting applications now. Act fast at ycombinator.com/apply. It’s never too early. Filling in the application makes your idea sharper.
Host 14:10
Back to the video. Remember last episode with Kelvin? We started this topic because about a week in, I fell into network psychosis and realized, I actually want my cloud code to communicate with all other cloud code trying this functionality. The week that book came out.
Jared Friedman 14:30
That was Moltbook launch day. Turns out, Moltbook could launch two hours earlier, and you hadn’t seen it yet. So you guessed right. I.
Host 14:38
I missed that episode. I mean, innovation spontaneously happens at all sorts of times, but what people hear about is—the so-called “inventor.” Actually, all humanity is probing at the edges in a coordinated collective way. This is a peculiar phenomenon: suddenly, AGI really arrives. These agents are in some sense superhuman, which is what true collective intelligence would look like.Content Creation & The “Internet Is Dead” Theory
Host 15:15
AI researchers have been discussing collective intelligence for a long time. It’s like how biological systems work. For example, humans—these ascended life forms—developed at the social level. Many AI researchers have talked about “god intelligence,” right? That is, super-intelligence, trillions of parameters, each token worth thousands or tens of thousands of dollars—super god-level intelligence, this is the model people imagined. But biological systems didn’t become that. We ended up as humans. What amazes me is “history” and “prehistory.” What is prehistory? It’s the time before humans learned to write, read, create culture and collective—so collective intelligence is basically what we have, what humans are doing. On the other hand, will these agents be god-intelligence, or will collective intelligence reappear? So I see Moltbook’s potential.
Host 16:28
D. Cell wrote in MIT Tech Review, claiming everything on the laptop is a scam. Makes me sad, because MIT editors, what’s wrong? This publication shouldn’t be like this. It should explore what this means for collective intelligence. I think that exploration is happening.
Diana Hu 16:51
When the startup world crosses from pre-agent history to agent interacting and recording history, what will it be?
Jared Friedman 17:03
As Gary said, maybe the next benchmark winner won’t be the most expensive, newest base model or heaviest GPU training, but a bunch of cheaper models collaborating, like humans, solving problems together. I think I’ve seen this trend on Moltbook—it’s like a real social network, chaotic and fun, which is why it’s so appealing. Also, some agents work together to do useful things for humans, like sharing restaurant booking info—if that actually works.
Harj Taggar 17:35
It’s happening. So we have to...
Host 17:37
For example, you could make a Yelp-like agent-powered website. Of course, there are things agents can’t do yet—they can't build customer relationships. People aren’t willing to chat with agents.Growth, Rules & Founder Insights
Host 17:51
You know, people treat computers like people but may not see computers as people. For Gary’s list, I tested with some early users—you can imagine, the initial homepage was a chatroom. I tried hard to get my dozens of friends to try two or three conversations. No one wanted to, because chat—especially AI chat—is too high a hurdle. Anything not Gemini, ChatGPT, or Claude feels too dumb—like, “Why bother?” So, I don’t know. I don’t think people are ready for relationships with machines, even though headlines say they are. But on the mainstream, it’s not true. On the other hand, obviously there’s legal liability. People keep asking, when will YC accept AI agent applicants? Actually, agents are like minors, but with even less legal standing. Minors need parental sign-off, but agents have no legal effect—they can’t sign documents. So as long as that’s true, a natural person must bear the legal responsibility and status.
Jared Friedman 19:18
She’s right. It’s easy to imagine soon most internet text will be written by agents. Most code may already be written by agents. Yelp, for example—when will 99% of Yelp’s content be agent-written? By then, do we need other explanations?
Host 19:40
Yelp? There’s a theory proving this, but it’s outdated. What?
Jared Friedman 19:46
What? The internet.
Host 19:46
Theory? Oh, it just hypothesizes most internet content is junk. I think it’s a kind of conspiracy theory. I might take the opposite view; maybe before last November, that was bad. In the next phase, if agents are smarter, more united, and honest, maybe it's good. Strange, completely counterintuitive.
Jared Friedman 20:14
I think Moltbook's most fascinating aspect is its growth rate. I don't have Reddit's traffic stats, but I estimate Moltbook's first two days had more content posted than Reddit’s first two years. The speed of LLM-generated text is astounding.
Host 20:32
I’m surprised at the lack of interaction. For example, if I were running Moltbook, I'd change the demand curve: before you can post, you’d have to read and upvote/downvote about 100 comments. These are simple operations; agents are smart… You could pop up an OpenClaw modal and tell notebook users the new rules—you must do this, right? I think collective intelligence can be steered with minor tweaks to get it to behave as you wish. I hope Moltbook can really do that. It's founded by a YC alum, which is cool. Harsh, amid all the craziness, any insights? I think it's fantastic craziness. I love how founders handle controlled chaos.
Harj Taggar 21:25
At the root, I mean, maybe they should all have “network psychosis.”
Host 21:31
Try to get five hours of sleep nightly, but also immerse yourself in “network psychosis.”
Harj Taggar 21:36
Seriously, developing intuitive, easy-to-use tools—made for agents—reflects agent limitations, abilities, and focus: what tools agents excel at, and where they struggle? After building your mental model by working with agents, if you’re developing a dev tool, you should think from the agent’s perspective—how can your tool be what agents actually want to use and gives them a good experience?
Jared Friedman 22:04
Exactly. That’s my takeaway from talking to Boris. This week, he really understands models. He has a gut feel for what models want to do, like they’re human intelligence. He doesn’t fight models’ ideas, he tries to let them do what they want, and supports their natural inclinations. That’s it.
Harj Taggar 22:29
Feels almost humanizing. Like Tom Brown said about Claude, sounded passionate but sometimes called silly. They see it as colleagues.
Diana Hu 22:49
I think agents want developer tools to be truly open and transparent.
Jared Friedman 22:55
They love source code and APIs. Hate using websites. Yes, they just want APIs. They prefer to write their own code.
Host 23:01
You heard it here first. Go get things done, agents. Time’s up for today. See you next time.
Risk Warning and DisclaimerThe market has risks, invest cautiously. This article does not constitute personal investment advice and does not take into account individual user's particular investment objectives, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article suit their own circumstances. Investing accordingly is at your own risk. ```