The investor who'd been attending Y Combinator demo days since the early 2010s paused mid-scroll through the Spring 2025 batch list. Something felt different this time. Not just the addition of a new seasonal cohort, though that was noteworthy enough. It was the density of a single idea, repeated with variations across nearly half the companies taking the stage that June afternoon in San Francisco.
Approximately 67 to 70 of the 144 startups were building AI agents.
Not AI features. Not machine learning tools. Agents—autonomous software meant to execute tasks, make decisions, and navigate systems without someone steering every move. The kind of technology that had seemed mostly theoretical a year earlier was now the premise for roughly 46% to 49% of one of Silicon Valley's most influential accelerator cohorts.
Even for YC, which has turned founder conviction into something of a house style, the concentration was striking.
Four Cohorts, One Big Bet
The Spring batch was part of an expansion. Y Combinator had announced in January 2025 that it was expanding from two annual cohorts to four, adding Spring and (eventually) Fall to the traditional Winter and Summer cycles. The move reflected both the accelerator's ambitions and the sheer volume of applications it was fielding. Startups, it seemed, weren't waiting for convenient calendar windows to get going.
Spring 2025 ran the standard three months, April through June, in San Francisco. The 144 companies each received $500,000: $125,000 for 7% equity via a post-money SAFE, plus an additional $375,000 through an uncapped MFN SAFE. By mid-June, roughly 1,500 investors and reporters had gathered for Demo Day.
TechCrunch's dispatch from the event captured the mood: "Nearly every presenting startup had something to do with AI—either building agents or tools for them." The observation wasn't hyperbole. If you were trying to spot a pattern in where young founders thought the puck was headed, you didn't need a flowchart.
The Categories Within the Category
What the 70 agent companies were actually building varied more than the headline number suggested. Some were constructing platforms—scaffolding for others to deploy their own agents. Sim, founded by Emir Karabeg and Waleed Latif, went the open-source route, offering tools for agent workflow creation. Prism promised deployment through a single API call. Naïve described itself as providing "unified agent infrastructure" for what it termed "AI employees that actually execute," a phrase that hovered somewhere between ambition and marketing.
Others focused on what you might call the unglamorous but necessary layer: infrastructure. Cua built cloud desktops specifically for agents, tackling the thorny problem of giving software autonomous access to computing environments without creating security nightmares. Airweave offered an open-source context retrieval layer. Capacitive went after data gateways—managing the complex dance of context, permissions, and security when agents need to touch enterprise systems.
Then there were the vertical plays. Den styled itself as "Cursor for knowledge workers," invoking the AI coding assistant that had become a reference point for what augmentation could look like. Eloquent AI, led by Tugce Bulut, built an operator for financial services, targeting task automation for certain workflows, though proving performance at scale would be the real test. Docket aimed at web testing agents, its founding team pulling from stints at Stripe, Citadel, Brex, and Patreon—the kind of pedigree that tends to open investor calls.
The "Cursor for X" framing showed up enough to become a pattern. If Cursor had demonstrated that AI could reshape how developers write code, Spring 2025 founders were searching for equivalent breakthroughs in legal work, finance, testing, analytics. Whether those domains would prove as receptive was an open question.
Early Wins and Quantum Bets
Not all 70 agent startups would make it, naturally. But a few signaled early momentum. Sim closed a $7 million Series A announced on November 12, 2025—five months post-Demo Day—validating the notion that open-source platforms could build sustainable businesses while maintaining developer flexibility. Eloquent AI raised $7.4 million announced around early September, with Business Insider profiling its approach to financial automation. The speed of both raises suggested certain investors were eager to place bets on vertical-specific agent plays with clear ROI narratives.
But perhaps the most unusual trajectory belonged to Sygaldry Technologies. Founded by Chad Rigetti, who'd previously led the quantum computing company that bore his name, Sygaldry wasn't building agents at all. It was building quantum-accelerated AI servers—infrastructure for the infrastructure, in a sense. On April 14, 2026, less than a year after Demo Day, Sygaldry announced a $139 million raise. Fortune covered it exclusively, noting Rigetti's wager that quantum hardware would become critical for scaling AI workloads as agents multiplied.
Whether that bet pays off is a longer-term question. But it illustrated how the agent wave was already spawning second-order bets on what would be needed if agents actually worked at scale.
The Backdrop
Spring 2025's agent concentration didn't materialize from nowhere. YC's Winter 2025 cohort, which had demo day'd in March, showed heavy AI representation as well. Partner Jared Friedman noted that roughly a quarter of the W25 batch had codebases that were 95% AI-generated—what he called "vibe coding," a term that captured both the speed and the slightly unnerving ease with which founders were shipping product. Increasingly, startups were building with AI and building for AI simultaneously.
CB Insights later mapped the Spring 2025 cohort across 18 categories within the agent umbrella: workflow automation, customer support, data analysis, and more. Alumni Ventures, in an investor recap published June 18, flagged additional themes: voice agents had a noticeable presence, and robotics—a category that had gone quiet in recent YC cohorts—showed signs of life again.
The diversity within the agent category mattered, perhaps more than the raw number. These weren't 70 teams chasing an identical wedge. Some pursued horizontal infrastructure, others vertical solutions. Some targeted consumers, others enterprises. The unifying thread was autonomy: software you could hand a goal and let figure out the path.
What It Meant
So what did Spring 2025 reveal, beyond the obvious fact that a lot of founders thought agents were the next big thing?
For one, the technical barriers to building them had dropped fast. The proliferation of foundation models, improved tool-use capabilities, and maturing frameworks meant small teams could ship agentic products in months rather than years. The infrastructure was there. The question shifted from "Can we build this?" to "What should we build, and who will pay for it?"
It also revealed founder conviction in a particularly pure form. Y Combinator doesn't tell applicants what to build—it selects from what shows up. That 70 teams independently landed on agents as their wedge suggests a shared intuition about where near-term value would concentrate. Right or wrong, that's a signal worth noting.
Whether the intuition proves correct is still being written. Some of these companies will scale. Many won't. That's how cohorts work. But Spring 2025 captured a moment when agents moved from research demonstrations to founder ambition, when the conversation stopped being about feasibility and became about application.
A year on, the concentration looks less like undifferentiated hype and more like the messy early phase of platform adoption—crowded, chaotic, generative. The agents are being built. Now comes the harder part: figuring out which ones the market actually needs.
