What’s popping in YC S25?

A roundup of the biggest themes from YC’s Summer 2025 batch

Every YC batch has a “feel.” Winter 2024 was the “AI everywhere” batch. Summer 2025 is the “agents batch.” It’s the first time the pitch is explicit: don’t hire a person, hire compute.

We’ve gone from copilots to employees. Last year, it was “help you write the email”. Now, it’s “answer the phone, collect the money, move the order into the ERP.” The common thread is closing loops end-to-end. That changes what matters. As agents step into roles that humans used to fill, trust, auditability, and auditability move front and center.

Other themes show up too: regulated-workflow AI (tax, legal, clinical), creator-economy infra (TikTok Shop, affiliate rails), embedded/firmware devtools, and geo-first wedges (LatAm, Pakistan). Consumer plays are surprisingly thin, with only a couple of creative tools and one social app.

What’s popping?

Agentic “employees” for ops P&Ls

This are the most literal version of “agents”: software that sits in a seat a human used to fill. A lot of companies are going after ops-heavy workflows - collections, order fulfillment, dealership leads. The pitch has shifted from copilots to closers.

• Altur – debt collection calls in LatAm.

• Acrely – voice AI concierge for HVAC service bookings.

• Wayline – front-desk assistant for property managers.

These companies are a test of trust. Budgets already exist for these roles, so on paper the swap is easy. The challenge is psychological: will businesses trust an agent to move money or handle customers without a human double-check? If that happens, “revenue per employee” stops being meaningful. We might end up with a new metric: revenue per dollar of compute. Headcount has always been the cap on growth. The ceiling shifts to compute, capital, and distribution.

There’s also a globalization angle. Call centers were the last arbitrage. Agents may be the next. If you’re already outsourcing collections to the Philippines, swapping to compute doesn’t feel like a leap.

Vertical AI for regulated fields

Think “Cursor, but for licensed pros.” These are industries where the bottleneck is people, not demand. CPAs, lawyers, and clinical trial coordinators are in short supply, and founders are dropping AI straight into the gap. AI feels like a natural fit anywhere the supply of talent is constrained.

Here are some examples:

• Magnetic – AI prep for CPAs.

• Vulcan Technologies – AI legal/regulatory tooling.

• Phases – patient recruitment for clinical trials.

These are less flashy than phone agents, but likely more defensible. Once filings, compliance, or trial records flow through a tool, ripping it out feels too risky. The moat isn’t cost savings, it’s the liability of leaving.

Agent-native infra

Most of S25 is applications, but there’s a thin (and important) layer of infra: tools that let agents interact with the world the way humans do. These aren’t sexy to end-users, but they’re critical scaffolding if agents are going to scale beyond demos.

Here are some examples:

• AgentMail – an email API that gives agents their own inboxes. Instead of bolting onto Gmail, it’s Gmail-but-for-agents, letting them send, receive, and process email natively.

• mcp-use – infrastructure for building and deploying custom Model Context Protocol (MCP) agents.

• DeepAware – optimizing data center ops with RL (AI optimizing AI infra).

Infra isn’t just plumbing, but it’s also distribution. Whoever owns the rails that agents use to talk to the outside world, email, WhatsApp, ERP APIs, will hold power, the way Twilio did for SMS.

If agents are going to live in the labor market, they’ll need the same “office tools” humans do: email, calendars, CRMs. AgentMail is the first clear stab at that.

Fintech agents

Finance are agents’ toughest proving ground.

Here are some examples:

• Finto – AI accounting for enterprise teams.

• Veritus Agent – AI voice agents for loan servicing and collections.

• Gaus – personalized AI agent for retail investors, scanning market data for trade ideas.

Finance can get adversarial - fraud, bad actors, regulators. Agents don’t just need to follow the rules, they need to defend themselves. If ops agents are the midterm, fintech agents are the final exam.

AI for sales/commerce enablement

This one feels more market-pulled than YC-pulled. TikTok Shop and affiliate rails are blowing up, and AI makes long-tail channels viable that humans wouldn’t bother with.

Here are some examples:

• Reacher – automated creator/affiliate outreach for TikTok Shop.

• Channel3 – API to list/sell any product online, with affiliate infra built in.

Feels like early infrastructure for the coming wave of AI-native storefronts. Also, distribution channels are mutating faster than companies can staff for. By the time you’ve built a human team, the platform has already shifted. AI is better suited for that pace. These startups aren’t just about saving cost; they’re about keeping pace with attention shifts on the internet.

Cursor for hardware

There’s a small but interesting cluster around devtools for hardware/embedded engineers. These companies line up with YC’s taste for technical founders scratching a personal itch.

Here are some examples:

• Embedder – assistants for embedded engineers.

• BootLoop – AI that writes/tests/debugs firmware on real hardware.

Hardware cycles has always been long and painful. Now AI might actually make it possible to shrink them. This isn’t just about making engineers happier. If firmware cycles get shorter, whole downstream industries like robotics or consumer electronics move faster.

AI-native security

If S25 is about putting agents into the loop, then security is about making sure the loop doesn’t snap. The first generation of security companies here aren’t doing generic SOC automation—they’re protecting agent workflows themselves: identity, data, and audit.

• Alter – zero-trust identity and access control, designed for agent-to-agent and agent-to-human systems.

• Hyprnote (cross-post from healthcare-adjacent AI) – open-source notetaker that runs fully on-device or on-prem, keeping sensitive conversations out of third-party servers

This cluster is small but telling. Every time you let an agent touch PII, finances, or legal filings, you’ve created a new attack surface. Just as cloud created Okta, Crowdstrike, and Datadog, agent adoption will birth its own security giants.

Geo-first AI wedges

Classic YC play: start where incumbents don’t bother. This batch has some AI-native products aimed at LatAm and South Asia, with WhatsApp and voice as the natural distribution rails.

Here are some examples:

• Uplift AI – voice AI for Pakistani languages.

• Altur – LatAm-first debt collection agents.

These aren’t clones of US products - they’re shaped by local infrastructure. Cheap labor means these markets already think in terms of efficiency, smartphones provide distribution, and payments rails like UPI and Pix make transactions seamless. That makes LatAm and South Asia unusually strong testbeds.

SMB & industrial ops automation

These are low glamour, high ROI companies - the kind of workflows where AI quietly makes money.

Here are some examples:

• Flai – auto-responds to dealership leads

• burnt – order-to-ERP automation for food distributors

The value-prop is simple: in low-margin industries, every dollar saved drops straight to the bottom line. These workflows already have budgets, the ROI is easy to prove, and buyers don’t need to be convinced. That makes adoption fast and churn low.

RL environments/agent training grounds

If agents are going to be everywhere, they need places to learn. Less visible, but very important.

Here’s an example:

• Idler – RL-driven training environments.

In hindsight, whoever defines the benchmarks often shapes the whole field.

Healthcare-adjacent AI

Not new drugs, just less paperwork. The healthcare startups in S25 are mostly focused on administrative bottlenecks: clinical trial recruitment, record aggregation, secure workflows.

Here are some examples:

• Phases – trial recruitment.

• Codes Health – compiles scattered patient records.

• Hyprnote – on-prem AI notetaker, relevant for HIPAA environments.

These products live in the messy middle: big demand, high stake (no tolerance for sloppy errors). Unlike ops agents, a mistake here isn’t just a lost lead, it’s a harmed patient. Adoption will be slower, but once embedded, these tools are far harder to rip out.

Consumer creative tools

Light consumer presence this batch - mostly creative tools, not new social apps.

• Riff: AI-assisted music production.

• RealRoots: guided friendship app for women

Predictions

Infra wave is coming.

Applications come first, infra comes second. Once a company is running more than a handful of agents, problems crop up: conflicting instructions, hallucinated actions, no audit trail. Expect S26 to tilt hard infra.

Geo-first will spread.

LatAm and South Asia aren’t just “cheap labor markets.” They combine low labor costs, high smartphone penetration, and modern payments rails. That makes them stronger sandboxes for agents than the US - and what works there will expand globally.

More infra around trust & compliance.

Not just evals, but also logging, audit trails, and liability insurance. If agents are doing work that matters, someone has to take responsibility. That creates demand for an entire compliance stack: audit logs, certifications, even insurance. Think SOC2, but for agents.

Consumer is overdue.

Consumer apps are almost missing in S25. The breakthrough probably won’t be a new social network. More likely, it’ll be agents embedded into existing platforms - TikTok, WhatsApp, Discord - where distribution is already solved. The consumer layer will sneak in sideways.

---

I invest at Weekend Fund. If you’re building in these spaces, or working on something that doesn’t fit neatly into these buckets, I’d love to hear from you - vedika@weekend.fund.

---