Nearly half. That's the figure that caught attention when Y Combinator unveiled its Spring 2025 Demo Day cohort in San Francisco this past June: 70 of 144 startups building what the industry now calls "agentic AI"—software that doesn't just assist but acts.
The concentration was unprecedented for YC, which has funded thousands of companies since 2005 but never seen a single technology thesis dominate a batch quite like this. Then again, the Spring cohort itself was an experiment. Announced in January, it represented YC's push to quadruple its cadence, adding Spring and Fall batches to the traditional Winter and Summer cycles. The reasoning? AI moves too fast for old rhythms.
Whether that gamble pays off remains an open question. But the roughly 1,500 investors who gathered for Demo Day that afternoon witnessed something like a collective hypothesis: that the next chapter of enterprise software won't merely augment human work—it will execute it.
When Software Stops Suggesting and Starts Doing
The distinction matters more than it might seem. Most AI tools launched over the past few years have operated as sophisticated advisors—drafting emails, summarizing documents, offering next-best actions. The Spring cohort skewed toward something bolder: systems built to complete entire workflows with minimal human intervention.
Take Aegis. Founded by Krishang Todi, Aarav Bajaj, and Dhanya Shah, the startup deploys AI agents to appeal denied health insurance claims—a process that typically involves navigating byzantine documentation, tracking down medical codes, and filing forms across multiple portals. The agent doesn't suggest how to appeal. It files the appeal.
Or consider Chestnut, which claims its agents automate 99% of mortgage lending work and has already secured licenses in Texas and Colorado. Approval AI takes a similar approach, building what it calls a mortgage copilot to handle origination tasks that usually require teams of loan officers. These aren't incremental productivity gains. They're architectural rewrites of how lending gets done.
Casco—whose team came from AWS's GenAI group—turned the lens inward, building autonomous security testing for AI systems themselves. A meta-problem, perhaps, but one that matters as agents proliferate. Mbodi AI went even further: its agents teach robots new skills through natural language, collapsing what used to require weeks of programming into something closer to conversation.
The breadth was startling. Willow built voice dictation that learns to write in each user's tone. Tegore applied agents to math tutoring. Zeon Systems brought AI-powered robotics to lab automation. The common thread wasn't the industry vertical but the core bet: that software could—and should—operate independently.
Plumbing the Agent Economy
Beneath the use cases, another layer emerged. Infrastructure.
As agents moved from novelty to assumption, founders began building the connective tissue that autonomous systems would need. Sim launched an open-source platform for agent workflows and claimed more than 70,000 developers were already using it—a figure that remains unverified by independent sources. Capacitive positioned itself as a "data gateway," enabling agents to pull from Slack, Notion, Google Drive, and Jira without the manual integration work that typically bogs down enterprise software deployments.
StarSling created what it described as an "agentic developer portal" for DevOps, orchestrating multiple agents across incident response and deployment pipelines. Plexe tackled predictive ML modeling with open-source agents. These startups were, in effect, betting that as agents become table stakes, teams will need better ways to connect, monitor, and—perhaps most critically—control autonomous processes at scale.
The infrastructure play made sense. If agents become ubiquitous, someone needs to build the pipes.
Four Batches, Faster Cycles, Higher Stakes
YC's decision to expand to four annual cohorts wasn't arbitrary. The accelerator framed the move, announced on January 20, 2025, as a response to AI's accelerating pace—a way to stay relevant when technology cycles compress from years to quarters.
The Spring batch ran April through June on YC's San Francisco campus. Each company received the standard $500,000 investment: $125,000 for 7% equity via a post-money SAFE, plus $375,000 on an uncapped most-favored-nation SAFE. Three months of small partner groups, weekly meetings, and alumni network access before the culminating Demo Day pitch.
Whether the faster cadence succeeds depends partly on how Spring cohort companies perform. Early indicators were... mixed. TechCrunch reported in March that a quarter of YC's Winter batch had codebases that were roughly 95% AI-generated—a striking figure if true. CNBC noted that same batch was growing revenue about 10% week-over-week, which would represent the fastest pace in YC's history. But those numbers came with caveats: small sample sizes, early-stage volatility, and the reality that most startups fail regardless of their growth metrics at month three.
The Spring cohort carried similar momentum, though with a sharper focus on autonomous execution rather than AI-assisted development. The distinction might prove meaningful. Or it might be a semantic difference that collapses once these companies scale.
Money Follows Conviction (For Now)
Post-Demo Day funding activity offered some signals. Lyra, which built what it calls an AI-native meeting platform, raised a $6 million seed round by late July, according to Business Insider. Mbodi AI secured investment in mid-June, though the amount wasn't disclosed. These weren't outliers—investor appetite for agentic startups appeared strong enough that YC hosted Agent Jam, a dedicated hackathon for AI agents, in November.
But early funding doesn't guarantee long-term success. The companies are young. The technology remains unproven at enterprise scale. And regulatory questions around autonomous systems—particularly in sensitive domains like healthcare and finance—loom larger than founders might acknowledge on stage.
Still, the Spring batch offered something like a referendum. A snapshot of where founders and capital were converging. Agentic AI had moved from fringe experiment to default assumption—at least within the specific bubble of Y Combinator's San Francisco campus that June afternoon.
The Question Nobody Asked on Stage
What nobody discussed during Demo Day, of course, was whether this many agent-focused startups in a single cohort represented opportunity or overcrowding.
Seventy companies building autonomous AI. Dozens more in adjacent infrastructure. All competing for attention from the same pool of investors, many of whom were already backing agent startups from previous batches. The math gets uncomfortable quickly: even if the agent thesis proves correct, most of these companies will fail. Some will pivot. Others will run out of runway before finding product-market fit.
That's startup math, admittedly. But the concentration added pressure. The companies pitching that afternoon weren't asking whether agents would work—they were competing to define which industries get automated first, and who gets to own that automation. It's a high-stakes game of musical chairs, and the music is still playing.
For now, anyway.
