Y Combinator Spring 2026 Startups: YC Spring ’26 Batch Overview

The Spring 2026 batch of Y Combinator is the most agent-heavy cohort YC has produced yet. But the interesting part is not that everyone is building agents. It’s what everyone has decided agents will need next.

Below is a full breakdown of the trends we’re seeing across all 194 launched companies, followed by a complete reference table so you can scan every startup in the batch by category.

Our Key Takeaways:

• B2B is about 61% of the batch (119 of 194 companies), and it leans heavily toward infrastructure and tooling rather than end-user apps.
• The real story is the agent supply chain. Dozens of startups are selling phone numbers, payments, identity, memory, sandboxes, and even insurance built specifically for AI agents, treating the agent as a new kind of customer.
• Industrials is the quiet heavyweight at roughly 12% (23 companies), powered by a deep defence cluster and robotics aimed at narrow verticals.
• Prediction markets showed up as a genuine micro-sector for the first time, with a handful of teams building the brokerage and derivatives plumbing around them.
• Geography is overwhelmingly San Francisco. The non-SF teams are few, and they cluster around hard tech, regulated markets, and a small New York finance pocket.

Complete List of All YC Spring 2026 Startups

Below is every launched company in the batch with a one-line description of what it’s building, its primary category, and its headquarters.

---

The Trends Shaping YC Spring ’26

We read every one-liner in the batch so you don’t have to. Here is where the founders are actually placing their bets.

Everyone is now an arms dealer to the agents

The clearest signal in this batch is not the agents themselves. It’s the explosion of companies treating “the agent” as a new economic actor that needs its own bank, lawyer, and phone line. AgentPhone is selling agents phone numbers. Allowance gives them scoped payment credentials, and Klaimee will sell them liability insurance. Clawvisor handles their authorization, Memory Store and Wato hold their shared memory, Ardent rents them database sandboxes, and StableBrowse gives them a browser. primitive is even building communication rails between agents. The most telling pair: RentAHuman is a marketplace for agents to hire humans, and Humwork connects agents to human experts in 30 seconds. When the new buyer in your market is software, an entire services economy springs up to sell to it, and that economy is most of this batch.

Almost everyone is selling a “layer,” not a product

Read enough of these pitches and a pattern jumps out: nobody wants to be the app, everyone wants to be the plumbing. You’ll find an authorization layer (Clawvisor), an intelligence layer for drones (Maquoketa), for drug manufacturing (AbInitio Bio), and for manufacturing generally (Korso), a brand layer for commerce (Kinect), a browser layer (StableBrowse), a runtime layer (ReasonBlocks), a context layer for sales (Salesgraph), a derivative layer for prediction markets (Totalis), and an infra layer for agent experience (Scope). Add the wave of “infrastructure for” and “OS for” framings and the conclusion is hard to miss. The batch has collectively decided the durable money sits underneath the agents, in the picks and shovels, not in the faucet.

Prediction markets quietly became a sub-sector

A category that barely registered in past batches now has its own little ecosystem. River Markets is building prime brokerage for prediction markets, Oddpool is doing institutional infrastructure, Totalis is building a derivatives layer, and ValCtrl is pitching a world model for what it calls “Prediction Finance.” Around the edges sit adjacent retail-trading plays like Mochatrade (US stock perps for Indian traders). One micro-sector with brokerage, derivatives, and a world model is no longer a novelty trade. It’s an emerging market that founders expect to need real financial infrastructure.

Coding agents grew up and need a workplace

Last year’s coding copilots have produced a full tooling layer around them. There are IDEs built specifically for the agent era (Superset, Zenbu, Ara), harness runners for the cloud (Replicas, Limrun), sandboxes for running them safely (Runtime), code review designed for mixed human and agent teams (Stage), and a thick observability stack to watch what they’re doing (Sazabi, Superlog, Armature, BentoLabs). The quiet assumption underneath all of it is that the “developer” using these tools may itself be an agent, which flips the entire developer-tools playbook on its head.

The pitch is “rip out the incumbent,” not “beat the incumbent”

Founders are no longer positioning against feature gaps. They’re naming the giant they intend to replace and rebuilding it from scratch as agents. Modern is “AI-native ServiceNow,” Netter is “AI-native Palantir for mid-market,” Soria is an “AI Bloomberg Terminal,” Smartbase and qomplement are AI-native ERPs, and Mount is pitching itself as an actual AI insurance carrier rather than a tool that sells to one. The “Vanta for X” naming convention has fully arrived too, with Complir (Vanta for physical products) and Fuchsia (Vanta for hardware). The bet is that incumbents built around human workflows can be unbundled entirely, not merely out-featured.

Physical AI got specific

Robotics in this batch is barely “general purpose.” It went vertical. You’ll see machine intelligence for welding (Advanced Metal Research), robots and software for farms (Synphony), humanoid robots for elder care (Twolabs), robotic warehouse staff (InLoop Robotics), robotic pit-stops for self-driving cars (Aseon Labs), and surveillance “robot birds” (Ornadyne). Underneath them sits an equally specific supply chain for robot training data and world models (shotwell.ai, Hub.xyz, Human-data plays, Intelligence Factory). Defense is the densest hard-tech cluster of all, spanning attack drones (Tenet), counter-drone systems (Surtr, Arlo), submarine drones (AICE), and mission systems (9 Mothers), with Apollo Atomics building compact nuclear machines on the edge of it.

The Bottom Line

The Spring 2026 batch reads less like a list of products and more like the early scaffolding of an agent economy. The headline isn’t that AI is everywhere, that’s been true for a few batches now. It’s that founders have stopped building agents and started building the world agents will live in: the identity, payments, memory, sandboxes, and insurance they’ll need, plus a fresh attempt to rebuild every enterprise incumbent from scratch.

Whether that thesis holds is the open question. A market where the biggest customer is other software is exciting and fragile in equal measure, and a lot of these “layer” companies are betting the same scaffolding will still be standing in three years. But if YC’s history is any guide, several of the 194 below will matter a great deal. Here’s the full list so you can make your own call.

---

Y Combinator has spent two decades setting the pace for early-stage tech. Since 2005 it has backed more than 5,000 companies with a combined valuation north of $1 trillion, roughly 6.5% of them have reached unicorn status, and a quarter of those unicorns have crossed $10 billion. Founders who go through the program are about 45% more likely to raise a Series A than those who don’t. That track record is exactly why each new batch gets read like a tea leaf for the next few years of technology.

• See Y Combinator Winter 2026 Startups List

---

Spring ’26 launches are still being added by Y Combinator, so this list reflects the 194 companies launched at the time of writing.

If you’re after software deals for founders, take a look at the deals offered through YC Startup School.