What Y Combinator’s Latest Batch Reveals About The Future

Building The Infrastructure For AI Agents

Building an AI agent is becoming easier. Deploying one inside a business is not.

The gap between a successful demo and a reliable production system has created an entirely new category of startups. The companies in this batch are building the memory, communication, validation and observability layers that AI agents need to operate in the real world.

ReasonBlocks is tackling one of the biggest obstacles facing enterprise AI adoption: cost and reliability. Rather than building another model, the company is focused on making existing models practical enough for real-world deployment. Its platform stores successful reasoning patterns from previous runs and injects them into future workflows, helping agents avoid repeating mistakes while significantly reducing token consumption. On SWE-Bench Pro, ReasonBlocks reports a 52% reduction in token usage alongside a 42% improvement in accuracy using the same underlying model.

“AI agents in production are expensive and often unreliable, with companies now paying six figures a month for systems that still fail too often to trust,” says Sajeev Magesh, cofounder and CEO.

Runtime is focused on one of the least glamorous but most important challenges in AI adoption: getting agents into production. While much of the industry remains focused on model performance, the company is building the infrastructure required to deploy, manage and operate autonomous systems at scale.

The founders believe the next generation of AI winners will not be the companies building models, but the companies making those models reliable enough to run real businesses. As organizations move from experimentation to deployment, Runtime is betting that operational infrastructure—not model intelligence—will become the bottleneck.

Memory Store cofounders Ishita Jindal and Diwank Singh met in 2018 through a shared obsession with the movie Her and later set out to build AI assistants themselves. After building thousands of agents through their open-source platform Julep, they discovered that agents repeatedly forgot context and made the same mistakes. That realization led them to create Memory Store, a shared memory layer for both humans and AI agents.

“What separates one company from another stops being how well it executes,” says cofounder Ishita Jindal. “It becomes what the company knows that nobody else does.”

AgentPhone provides phone infrastructure designed specifically for AI agents. Its founders believe that if autonomous systems are going to operate alongside humans, they will need trusted identities and communication channels just as people do today.

“Every human has a phone number,” says cofounder Meet Modi. “It’s how the world identifies, reaches and trusts you. AI agents don’t have that yet.”

Meanwhile, Arga Labs is building what its founders describe as the validation layer for AI. The company creates digital twin environments where organizations can safely test AI agents before deploying them into production systems.

Sazabi is building an AI-native observability platform designed to automate incident detection, root-cause analysis and response.

“Monitoring is dead,” says Sherwood Callaway, a two-time YC founder, a16z scout and software engineer with more than a decade of experience. “The future is agentic alerts.”

Callaway is no stranger to the problem. Before founding Sazabi, he spent more than a decade building infrastructure systems, including observability platforms at Brex. The founding team includes early Brex engineers, former observability founders and infrastructure specialists, giving the company deep expertise in the category it hopes to reinvent.

Sazabi has also taken an unusual approach to fundraising. Rather than relying on a small group of institutional investors, the company has assembled a network of more than 100 angel investors, including founders and engineering leaders from companies such as Browserbase, LangChain, Graphite and Daytona. The strategy provides not only capital but also direct access to many of the builders shaping the next generation of AI developer tools and infrastructure.

Modern is applying the same AI-native thinking to enterprise service management. The founders describe the company as an AI-native successor to ServiceNow, built around a simple premise: enterprise software has become extremely good at tracking work, but not doing it.

“The software has always been a tracking system. Humans still do the actual work,” says cofounder Seb Poole.

Modern’s agents are designed to autonomously resolve service desk tickets while operating inside deterministic workflows that enterprises can audit and trust. Rather than adding AI to existing software, the company is rebuilding the platform around agents from the ground up. The founders believe the next generation of enterprise software will be measured not by how well it tracks work, but by how much work it completes on its own.

Reimagining Work

While much of the AI conversation focuses on productivity, several startups in the batch are attempting something more ambitious: replacing entire categories of operational work.

Dayjob emerged from an unlikely place. Its founders spent 18 months building software for waste-management companies before discovering that their customers were asking for something entirely different.

Every morning, transport planners spent hours manually building routes, only to watch those plans break down almost immediately as conditions changed throughout the day.

The founders realized the opportunity was not to build better software for planners but to build the planner itself.

“We stopped building software and started building AI workers,” says cofounder George Postlethwaite.

Today, Dayjob’s AI agents can rebuild complex logistics schedules in minutes, and the company now sees a much larger opportunity: automating operational roles across industrial logistics.

Revnu is pursuing a similarly ambitious vision in marketing.

The founders first experimented with AI while bootstrapping businesses from their university dorm room, building automated systems that handled everything from software development to customer support. Their most successful experiments involved growth and marketing.

“We are automating marketing,” says cofounder George Jefferson. “AI has automated software engineering. The next step is automating growth.”

The company is building systems that can run growth experiments, measure outcomes and optimize campaigns autonomously. Its long-term goal is to automate marketing and growth for businesses in the same way AI is beginning to automate coding.

AI Comes To The Physical World

Not every startup in the batch is focused on software.

Lumius is reimagining ultrasound technology through what its founders describe as a “3D camera for the body.” The company combines advances in AI, computing and medical imaging to make ultrasound easier to interpret and use, with potential applications ranging from bedside diagnostics to surgical robotics.

Avea Robotics is tackling another challenge in the physical world. As robots move into factories, warehouses and kitchens, even small failure rates can create significant costs. Avea enables humans to instantly intervene when robots encounter problems, reducing downtime and increasing reliability.

General Aviation is modernizing an industry that has changed surprisingly little over the past several decades. The company is building what its founders describe as a new air traffic control system by connecting aircraft directly to the internet.

The idea has become possible thanks to low-earth-orbit satellite networks such as Starlink, which now allow aircraft to communicate through digital messaging rather than relying primarily on legacy voice-radio systems.

“We are building a new air traffic control system by connecting aircraft to the internet,” says founder Ben Frank.

If successful, the company’s approach could fundamentally change how aircraft communicate and operate in increasingly crowded airspace.

Bringing AI Into The Enterprise

More than 70% of Fortune 500 companies still rely on legacy desktop applications without modern APIs. Replacing those systems is expensive, risky and often impossible. As a result, many AI deployments stall before they ever reach production.

Minicor enables AI agents to interact directly with those systems through intelligent desktop automation. By combining deterministic code with agentic workflows, the platform can create, monitor and repair automations without requiring enterprises to replace their existing infrastructure.

The company’s founders discovered that legacy software represented one of the largest barriers to enterprise AI adoption. During the batch, Minicor tripled monthly recurring revenue while sharpening its vision of becoming the infrastructure layer connecting AI to enterprise systems that may never receive modern APIs.

The opportunity highlights a recurring theme throughout the batch: many of the biggest AI opportunities are not about replacing existing systems but making them usable.

Rebuilding Trust

As AI becomes more capable, trust is emerging as one of the industry’s most valuable commodities.

Several founders in the batch are focused on solving exactly that problem.

Complir is bringing AI into product compliance, a category that remains heavily dependent on PDFs, spreadsheets and manual reviews.

“Compliance—the most boring word in commerce—is one of the most important infrastructure plays of the decade,” says Gustav Bang, cofounder and CEO of Complir.

Building the company from San Francisco during Y Combinator while serving a primarily European customer base has required unusual working hours. Bang says he and cofounder Tine Kühnel often high-five each other in the middle of the night as one finishes customer onboarding calls and the other begins the next round of meetings. With European customers operating eight to nine time zones ahead, the founders have spent much of the batch working through the night to stay close to users.

The company believes compliance will increasingly become a competitive advantage rather than a regulatory burden.

Klaimee is developing insurance products designed specifically for AI agents. Traditional insurance products were built to protect humans and computer systems, not autonomous software capable of making decisions on its own.

Trust is a theme that extends beyond the product itself. Cofounders Ines Boutemadja and Julien Catonnet have known each other for more than a decade, have been together for seven years and got married last year.

“People say your cofounder relationship is like a marriage,” says Julien. “In our case, it literally is.”

The founders believe the rise of AI agents will create entirely new categories of risk and require entirely new categories of insurance.

Kinro is building what it describes as an AI-native broker capable of helping small businesses navigate coverage decisions, access markets and manage policies over time.

“The next generation of financial help will start as a conversation, not a form,” says Pierre-Alexandre Kamienny, cofounder and CEO of Kinro.

Taken together, these companies suggest that trust infrastructure may become one of the most important markets of the AI era.

Solving AI’s Infrastructure Challenge

Behind every AI breakthrough sits an infrastructure challenge.

As demand for computing power accelerates, several startups are focused on making the underlying systems more efficient.

Expanse believes a significant portion of the world’s computing shortage already has a solution.

“Half of the world’s compute is being wasted,” says Ismaeel Bashir, cofounder and CEO.

After discovering millions of dollars in unused capacity inside a single data center, Bashir and his cofounders, Nikodem Bieniek, Yafet Melake and Eren Mendi, set out to build a platform that identifies idle capacity inside data centers and reallocates those resources to AI workloads.

ProjectX is building a new computing architecture designed for a world where humans and AI agents work simultaneously. Its founders believe modern operating systems were never designed for an AI-native future and that entirely new approaches to computing may be required.

Powering The AI Era

If compute is becoming scarce, energy may become even more important.

Apollo Atomics is betting that the future of AI will require a new generation of nuclear power. Founder Assil Halimi spent more than a decade working across the nuclear industry before reaching a conclusion that shaped the company’s strategy.

“The industry’s biggest obstacle is not physics, but implementation,” he says.

Rather than attempting to reinvent nuclear technology from scratch, Apollo is focused on commercializing compact reactors built on proven designs. The company believes that simplifying deployment rather than reinventing the science offers the fastest path toward expanding global energy capacity.

That approach appears to be resonating. During the YC batch, Apollo expanded from a single commercial agreement to more than 20 gigawatts of letters of intent and partnerships.

As AI drives unprecedented demand for electricity, startups like Apollo suggest that some of the most important AI companies of the next decade may not be software companies at all.

AI Is Reshaping Entire Industries

While many startups in the batch are building the infrastructure that powers AI, others are applying the technology to transform entire industries.

MochaTrade is bringing new financial products to global investors. The company gives Indian traders access to U.S. financial markets through perpetual futures infrastructure, addressing a longstanding gap in cross-border investing.

“We built MochaTrade because we were tired of being treated like second-class participants in the markets we actually wanted to trade,” says Utkarsh Sinha, cofounder and CEO.

KelAI is applying AI to institutional investing. Founder Jeremie Cohen spent six years at WorldQuant before launching the company, which aims to transform investment research into a continuously compounding intelligence system. Rather than replacing investors, KelAI is focused on accelerating how investment ideas are generated, tested and refined.

Asendia AI is targeting one of the world’s largest labor markets: recruiting. Founded by former engineers from Google, Microsoft and Infineon, the company believes AI can automate many of the repetitive workflows that currently consume recruiters’ time, allowing organizations to hire faster and more efficiently.

“We didn’t come from recruiting—we came in angry at a broken process,” says Rihab Lajmi, cofounder and CEO. “Turns out that’s a better starting point than being comfortable with it.”

Before building the product, the founders spent months shadowing recruiters and staffing firms, observing an industry still heavily dependent on spreadsheets, manual outreach and repetitive administrative work. Today, Asendia is building AI recruiters designed to automate sourcing, outreach and candidate management at scale.

As AI-generated content floods the internet, Manicule is betting that high-quality technical content and developer education will become more valuable, not less. Founded by 18-year-old Shreyans Jain and Naman Bansal, Manicule helps developer-focused startups build credibility and reach technical audiences in an increasingly noisy digital landscape.

Madrone is addressing a less visible but increasingly critical challenge facing the AI industry: cooling. As companies race to build ever-larger data centers, power—not chips—is becoming the primary constraint. The founders estimate that roughly 30% of a facility’s power can be consumed by cooling rather than computation.

Madrone develops cooling systems designed to reduce both power and water consumption in data centers, particularly in the hot, dry regions where many new AI facilities are being built. The opportunity emerged from a problem cofounder Erik Meike first began exploring six years ago, long before AI data centers became a global priority.

“We started this company because this is the biggest sustainability challenge we could be solving right now,” the founders say.

During the YC batch, Madrone increased its cooling capacity by 100x, accelerating plans to scale manufacturing and deploy the technology commercially.

What This Batch Reveals About The Future

Taken together, the companies emerging from Y Combinator’s latest batch reveal something important about where artificial intelligence is headed next.

For the past several years, the AI race has been defined by larger models, better benchmarks and increasingly capable systems. But the startups in this batch suggest a new phase is beginning.

The next challenge is not creating intelligence. It is deploying it.

As AI agents move into real businesses, they will require entirely new layers of infrastructure: memory systems, communication networks, observability tools, compliance platforms, insurance products, enterprise integrations, computing resources and energy generation.

At the same time, founders are applying AI to transform industries that have historically changed slowly, from logistics and recruiting to healthcare, finance, aviation and manufacturing.

The result is a broader shift in how technology is being built. Rather than creating standalone applications, many of these startups are developing the systems that will allow AI to operate reliably inside the real world.

Whether any individual company succeeds remains to be seen. But together, they offer a glimpse into the future: an economy increasingly powered by AI agents and the infrastructure required to support them.

If the first generation of AI startups focused on intelligence, this generation is focused on everything that comes next.