YC S25: The Complete Analysis

While individual companies may look like noise, cohorts reveal signal.

After analyzing all 153 companies in Y Combinator's Summer 2025 cohort, one thing became clear: this isn't just another cohort. It's a turning point.
This analysis covers the agentic workforce thesis, the infrastructure enabling it, the physical-world constraints that can't be coded away, and how these patterns are already shaping YC's next batch.

The Agentic Workforce: YC S25 and the Dawn of the Agentic Workforce

Across the 153 startups, more than 80% are AI-focused, and about a third are building autonomous agents.

In past cohorts, AI showed up as a feature - copilots, recommendation engines, automation layers. In S25**, AI is the product**.

That shift shows up in two ways I found especially interesting.

First, there's the rise of the Digital Coworker. Startups are creating AI agents that execute complete workflows, not just assist humans. We're seeing AI pharmacy techs (Pharmie AI), AI paralegals (Kalinda AI), and AI employees for real estate (Closera). They're all building systems that execute, not just software that assists - a quiet but meaningful distinction.

Second, the shift happening at the systems level. Founders are rebuilding core software as AI-native. Startups are re-architecting entire industries from the core outward, creating AI-native ERPs (Stockline), security platforms (Riverbank Security), and hardware compliance systems (Normal).

To me, this marks the real shift - the rise of companies where work itself becomes a service.

The Engine Room: The Infrastructure Powering the Agentic Era

As someone who spent much of my career on the infrastructure side, my first thought isn't about the agents themselves, but what it takes to run them reliably at scale. These AI-native systems won't just reshape workflows - they depend on a new foundation beneath them.

40+ startups in YC S25 are building something less flashy but more critical: the infrastructure determining whether autonomous AI succeeds or fails at scale.

I call this the Engine Room of AI.

Without this, even the most impressive applications become too expensive, too unreliable, or too dangerous to deploy.

Here's what's taking shape in the engine room:

1. The Base Layer: Specialized Models & Hardware

Specialized models, custom hardware, and the low-level code that ties them together form the foundation of the next AI wave.

• BootLoop.ai and Embedder: Automating firmware and embedded development
• Sigmantic AI: Accelerating silicon design itself
• Stellon Labs: Building "tiny frontier models" for edge devices

2. The Data & Memory Layer: Intelligence Depends on Context

AI systems are only as good as what they can remember and access.

• Epicenter and Rellings Systems: Persistent memory and world-model datasets
• Halluminate and MangoDesk: Training and evaluation pipelines
• Liva AI and Panels: Multimodal datasets for voice and video

3. The Compute Layer: The Economics Problem

The bottleneck isn't just intelligence - it's the staggering cost of running it. Whoever solves inference economics wins the infrastructure layer.

• DeepAware AI: Optimizing energy use in data centers
• Lilac: Unlocking idle GPUs through marketplace models
• Luminal: Compiling models for any GPU, slashing inference costs

4. The Ops Layer: DevOps for Agents

You can't deploy what you can't test, monitor, or fix.

• AgentHub: Simulation environments for testing agents before production
• Nixo: Ops platform giving FDEs instant visibility into customer workloads
• Nottelabs: Self-healing web agents that adapt to interface changes
• Traceroot.ai: AI agents that automatically fix production bugs

5. The Governance Layer: Building Trust in Autonomy

As systems gain autonomy, the trust layer becomes non-negotiable.

• Alter: Zero-trust identity and guardrails for autonomous systems
• Nuntius: Platform for making models follow specific rules
• Truthsystems: Real-time governance agents

There's a temptation to think of infrastructure as "plumbing" - important but not strategic. In AI, infrastructure doesn't just support the application layer. It defines it.

Compute costs determine which use cases are viable. Memory architectures determine what agents can remember. Governance frameworks determine what they're allowed to do.

The agents get the headlines. The infrastructure sets the boundaries of what's possible.

The New Industrial Frontier: Minerals, Drones, and the Geopolitics of Supply Chains

Eleven companies are building for a world where supply-chain resilience, strategic autonomy, and domestic industrial capacity become national priorities. They're solving problems in the physical world.

Three examples stand out:

• Albacore Inc. builds autonomous undersea vehicles for deterrence and maritime ISR missions. Founded by naval engineers and autonomy researchers bringing next-generation robotics to contested waters where reliability and range are non-negotiable.
• Duranium is reshoring titanium and magnesium production using a carbon-neutral process. Founded by a DoD supply-chain advisor and a molten-salt electrolysis expert tackling metals that underpin defense, aerospace, and advanced manufacturing.
• Verne Robotics is teaching industrial robots new skills in hours instead of months, collapsing the time needed to deploy automation. Founded by engineers from Columbia, Stanford, and Microsoft Azure Copilot solving the core bottleneck in industrial robotics: adaptation speed.

Others in this cluster: dScribe AI (computer-vision inventory), F4 (engineering drawings), Flywheel AI (autonomous excavators), Juxta (GPS alternative), Nexa Labs (cattle monitoring), Nox Metals (software-first metal distribution), Normal (hardware compliance automation), and Perseus Defense (counter-drone systems).

Software moves fast because iteration is cheap. AI is collapsing feature moats - when any feature can be replicated in hours, having the feature stops being a durable advantage.

Hardware moves slow because physics doesn't negotiate. But when hardware works, it's defensible in ways software isn't.

You can't fork a supply chain. You can't replicate a decade of metallurgy R&D in a weekend.

The Rebuilding Problem: What You Can't Retrofit in Regulated Industries

The cohort's regulated vertical plays - healthcare, finance, enterprise operations - reveal a pattern I didn't expect.

It's not about features. It's about what you can't retrofit.

Legacy platforms in regulated industries face three constraints that create a genuine dilemma:

• 📋 Compliance by Design: Audit trails built for human decision-making don't translate to autonomous agents. When regulators ask "who's liable when AI makes the mistake?", retrofitting an answer into existing systems is exceptionally difficult.
• 📊 Economic Model Inversion: Per-seat pricing assumes humans executing the workflows. When agents handle the work and humans shift to oversight, the unit economics flip. Transitioning pricing models while maintaining current revenue is a challenge most companies struggle to navigate.
• 🔧 Workflow Architecture: Platforms built for heavy human touch become agent-first only through architectural rewrites. Bolting autonomous workflows onto systems designed for manual processes rarely works.

And that creates an opening to build from zero.

Who's Building from Zero:

• Healthcare: b-12 (chemical synthesis copilot), CareSwift (AI scribe for ambulances), Wedge (governance layer for hospital AI)
• Finance: IronLedger (property accounting agents), Magnetic (vision model for tax prep), Socratix AI (fraud investigation coworkers)
• Enterprise: Certus AI (restaurant phone AI), Comena (order entry automation), Janet AI (AI-native project management)

These constraints aren't problems you solve with an AI feature release. They require architectural decisions most incumbents won't make.

Connecting S25 to F25 Cohort: The Patterns Amplify

The YC Fall 2025 cohort just started, and YC's Request for Startups (RFS) - their public list of problems they want founders to solve - shows some of the S25 patterns that are amplifying.

⚙️ Infrastructure: From Single Agents to Fleets

S25 had 40+ companies building infrastructure for agentic workflows - observability, optimization, evaluation, and memory.

The F25 RFS scales this: Infrastructure for Multi-Agent Systems - distributed workflows managing hundreds of thousands of subagents in parallel.

To me, this signals the shift from "can we make one agent work?" to "can we orchestrate thousands?" The infrastructure companies building in S25 are positioning for this next wave.

🏢 Enterprise: From Bolted-On to Built-In

The F25 RFS explicitly calls for rebuilding Salesforce and ServiceNow as AI-Native Enterprise Software - intelligence embedded throughout, not added as features.

This validates what S25's vertical SaaS plays revealed: incumbents face constraints they can't solve with feature releases. Compliance architectures, pricing models, and workflow designs all need rebuilding from zero.

The opening isn't in building better features. It's in building what can't be retrofitted.

💰 Team Size: From Headcount to Leverage

The "First 10-person, $100B Company" RFS makes explicit what S25 implied.

When AI handles execution and humans provide oversight, the leverage multiplier changes fundamentally. S25 showed the agentic workforce thesis. F25 is asking: how far can it go?

Single agents become fleets. Retrofitted features become rebuilt architectures. And the physical world - the thing software can't abstract away - reasserts itself as the ultimate constraint.

The Takeaway

We're still early in understanding what makes AI companies defensible. The pattern-matching from the last decade breaks. Unit economics flip when agents replace seats. Moats shift from network effects to things that exist in atoms or require regulatory trust.

The companies that matter will be the ones that understood: you can't fork physics, you can't retrofit trust, and leverage only matters if the foundation can support it.

What's Next

I'm tracking YC F25 as these patterns evolve - watching for what amplifies, what breaks, and what new constraints emerge. The agentic era isn't a single cohort. It's a multi-year story, and we're still in the opening chapters.

Connect with me on LinkedIn for commentary on startup patterns as they emerge.

Appendix: All 153 YC S25 Companies

The complete list of every company in Y Combinator's Summer 2025 batch. Also available as a spreadsheet.