The Demo Day traffic patterns told you something was up before a single pitch began. As roughly 1,500 investors filed into Y Combinator's Spring 2025 showcase in mid-June, the agenda listed over 140 startups. What wasn't listed: that somewhere close to half of them were building nearly identical businesses.
Not identical in the way VCs usually complain about—yet another food delivery app, another SaaS dashboard. This was different. Approximately 70 companies, by widely reported estimates, were building AI agents. Not chatbots. Not the LLM wrappers that flooded pitch decks a year earlier. Autonomous software designed to actually do things: hunt for security holes, run QA testing loops, book your meetings without you lifting a finger.
The concentration was stark enough that TechCrunch observed "nearly every presenting startup had something to do with AI"—either building agents outright or selling the infrastructure to support them. This wasn't a category gradually taking shape. This was a land rush, and everyone apparently got the memo at the same time.
Seventy Variations on a Thesis
Y Combinator doesn't release official breakdowns by sector, which means the tallies come from investors and analysts doing their own math. Alumni Ventures, dissecting the batch in a June 18 analysis, sorted the agent companies into nine subcategories. Nine. From QA automation to voice interfaces to something they called "agents for agents"—tools to build the tools.
The Spring cohort came as YC was transitioning to a new four-batch-per-year structure announced the previous September. The deal terms remained standard: $500,000 per company ($125,000 for 7% equity, another $375,000 via uncapped MFN SAFE). Three months from April through June, culminating in that packed room where half the companies were selling some version of the same dream.
What separated this from typical AI hype cycles—and there have been several by now—was specificity. These founders weren't waving their hands about "AI-powered solutions." They were building for narrow use cases that genuinely hadn't been possible twelve months earlier. The technology, they insisted, had finally caught up to the vision.
Maybe.
The Browser Becomes Programmable
Take the cluster focused on browser automation. Propolis, launched by Marc Papazian (Palantir alumnus) and Matt White (previously at Airtable), built what they termed "hands-off QA via intelligent browser agents." Synthetic users testing your software while you sleep. Asteroid pitched "browser agents as a service" aimed at regulated industries, claiming their approach ran "100x cheaper and more robust" than previous methods, thanks to recent computer-use models.
Hyperbrowser offered infrastructure—scalable, undetectable browsers with proxy rotation and CAPTCHA-solving baked in. BrowserOS went open-source, positioning itself as the foundational layer for automating any browser-based workflow.
Different companies, different angles. But the through line was unmistakable: computer-use models from Anthropic and others had crossed some threshold. What once required thousands of brittle, hand-coded scripts could now, allegedly, be handled by a single agent with the right scaffolding. Multiple founders told variations of this story. The infrastructure had matured. The moment, they believed, had arrived.
Whether investors fully bought it or were hedging their bets across dozens of similar pitches remained an open question.
Security Agents That Never Sleep
MindFort—founded by Brandon Veiseh, who'd spent time at ProjectDiscovery, alongside Akul Gupta and Sam Berston—built autonomous security agents for continuous penetration testing. Jazzberry promised AI-driven bug discovery with actual code execution, not just static analysis.
These weren't agents meant to replace developers. They were designed to probe systems at speeds and scales manual testing couldn't touch. According to several Demo Day attendees, this angle drew particular interest. Perhaps because the ROI argument felt clearer than broad automation promises. Companies already spend heavily on security testing. An agent running 24/7, catching vulnerabilities before they hit production? That's a line item executives understand.
Bluejay, which reported raising a $4 million seed in June from Floodgate, PeakXV, and YC, focused on trust and safety for AI interactions. The pitch centered less on raw capability than on reliability—making agents work consistently enough to trust them with production systems. A subtle but revealing distinction.
Selling Picks and Shovels
Not everyone was building agents. A meaningful subset was building for agents.
Den, which TechCrunch singled out as "one of the hottest companies" from the batch, positioned itself as "Cursor for enterprise knowledge workers." Sim Studios built an open-source agent builder with a Figma-like canvas for workflows. (The company reportedly acquired the domain sim.ai for $220,000, though the timing on that transaction remains somewhat unclear.)
The LLM Data Company focused on evaluation and reinforcement learning. Theta built what it called a "self-learning layer" to reduce the steps agents need to complete tasks. RunRL, founded by Andrew Gritsevskiy and Derik Kauffman, offered "reinforcement learning as a service" for real-time optimization.
These companies were making a different bet. Even if specific agent applications churned—and some inevitably would—the underlying need for training infrastructure, evaluation frameworks, and reliability layers would persist.
Airweave, from Lennert Jansen and a co-founder listed only as Rauf, built connectors to make workplace data searchable by agents. Cotera promised "the easiest way to put reliable AI agents into production." Atla pitched an improvement engine using evaluations to surface critical failures.
When Alumni Ventures categorized the batch, they created a dedicated bucket for "agents for agents." Infrastructure plays assuming agents would proliferate and focusing on making them actually deployable. Perhaps the smarter bet, if history is any guide.
What It Signals (and What It Doesn't)
The Spring batch followed YC's Winter 2025 cohort, which included 160 startups and, according to TechCrunch that March, "no shortage" of agent-focused companies. CNBC reported that YC President Garry Tan described the Winter batch as growing "10% per week" in aggregate—the fastest growth in the fund's history, driven heavily by AI momentum.
But Spring escalated. When nearly half a cohort converges on one category, you're looking at either collective delusion or genuine platform shift. The investor appetite in the room that day suggested the latter. Multiple companies announced seed rounds in subsequent weeks, though exact figures mostly come from jobs pages and founder posts—hardly the most rigorous sourcing.
The concentration also revealed live subcategory formation. Voice agents formed their own cluster: Atlog, Kanava AI, Lyra, SynthioLabs, Trapeze, VoiceOS, Willow. Healthcare got vertical plays like Tyran's "AI Health Agents for your app." Bear built something it called "Get recommended by AI Agents," betting that agent-driven distribution would become a channel unto itself. Vantedge AI created an agent marketplace for investment workflows, anticipating institutional adoption.
What made the batch notable wasn't just volume. It was the diversity of approaches within a single category. Not clones, but different layers and use cases of what they collectively believed would become a new platform.
Whether they're right is another matter entirely.
Too Early or Right on Time?
Y Combinator's track record includes Airbnb, Stripe, Coinbase, DoorDash, Instacart, Reddit. It also includes hundreds of companies that never made it past Series A. The Spring 2025 agent wave might look prescient in a few years. Or it might be remembered as the cohort that showed up before the technology was ready.
The skeptical case writes itself: computer-use models remain unreliable, agents hallucinate with alarming regularity, and most enterprises won't trust autonomous software with critical workflows anytime soon. Maybe not for years. The optimistic case points to rapid model improvement, specific domains where agents already outperform humans (at least on benchmarks), and infrastructure bets that assume scaling is inevitable.
What's harder to dismiss is the conviction. Seventy companies don't pile into one category by accident. They're responding to what they perceive as a technological unlock—models that can finally interact with software interfaces reliably enough to build businesses on top of them.
Whether that perception translates into billion-dollar outcomes or cautionary tales about timing and hype cycles won't be clear for some time. Markets render these verdicts slowly, then all at once.
For now, Y Combinator's Spring 2025 Demo Day stands as a remarkable snapshot of collective belief. Seventy startups, each wagering $500,000 in investor capital and years of their professional lives, on the proposition that AI agents represent the next platform shift. Half a cohort, one bet, 1,500 investors watching.
The answers, as always, will come later.
