This article is a translation of the startup request list published by YC Lab in the summer of 2026. The translation omits the hardware, space, and supply chain sections, keeping only the AI and software sections. If you’re interested, you can read the original at https://www.ycombinator.com/rfs
AI is no longer just a “feature” — it’s starting to become the “foundation.” We’re excited about a new wave of startups: they are rebuilding software, services, and chips, and pushing AI into the physical world. Many of these directions come directly from YC founders sharing the changes they’re seeing at the frontier.
AI for Low-Pesticide Agriculture
Author: Garry Tan
https://www.youtube.com/watch?v=uwe-5PB9p4Y
Modern agriculture runs on chemicals. That worked for a while, but the problems are now obvious. Pesticide residues are everywhere: in food, in water, in soil. People around the world are worried about the long-term health risks of glyphosate.
Meanwhile, nature is adapting. Weeds and pests evolve. What used to work gradually stops working. So farmers spray more. Costs go up, profits go down. And the pipeline for new chemicals? Slower and more expensive than ever.
Farmers are trapped in a vicious cycle: use more chemicals → diminishing returns → spend more money → take on more risk. But they can’t just stop. Once pests gain the upper hand, crops die. Food prices rise. People could go hungry. This is a matter of life and death.
For a long time, this looked like an unsolvable problem. Not anymore. Because several things have changed at the same time:
AI can now “see.” It can identify individual weeds and pests in real time. Sensors and cameras have become cheap enough to deploy everywhere. Robots can operate with precision: treating a single plant instead of blanket-spraying an entire field.
Biotech is catching up too: microbes, peptides, RNA-based solutions — these are no longer science fiction. They can replace entire classes of synthetic chemicals. We can engineer plants to defend themselves, suppress weeds, and reduce external inputs.
And of course, AI itself keeps accelerating. Real scientific breakthroughs will be amplified by AGI, and now is the moment to witness that.
This is a massive shift. Agriculture is one of the largest markets in the world. If you can cut costs and raise yields at the same time, adoption won’t be slow — it will be explosive.
The company that can cut pesticide use by 90% while helping farmers grow more food? That’s not just a good business — it’s a generational company.
If you’re working on this — if you’re building tools that help farmers produce more with fewer chemicals — we’d love to talk to you.
AI-Native Service Companies
Author: Gustaf Alströmer
https://www.youtube.com/watch?v=lsRnwM_ZdJY
AI models are improving extremely fast, and they can now do complex work far beyond engineering.
Historically, services evolved into SaaS software. More recently, they evolved into AI copilots: the kind of thing most startups built from 2023 to 2025 — tools that help people do their jobs better.
What we’re really excited about now is the next step: AI-native companies that don’t sell software — they sell “services.” Instead of giving you a tool so you can do the work yourself, they just do the work.
This matters for a simple reason: total spending on services is many times larger than spending on software. And a huge share of services is already outsourced, which makes them easier to replace with AI-native products.
Areas we’re especially interested in include:
- Insurance brokerage
- Accounting, tax, and audit
- Compliance
- Healthcare administration
If you’re building a company that does business by “replacing services” — not just improving them — we’d love to hear about your progress.
AI Personalized Medicine
Author: Ankit Gupta
https://www.youtube.com/watch?v=fzVuQ2sYEn8
Intelligent agents are bringing healthcare to a new level of personalization. Using frameworks like Claude Code — which connect agents to a set of tools — we can now analyze personalized health data, whether it’s diagnostic tests, genomic scans, EHR data (electronic health records), or wearable device data, to deliver highly accurate recommendations tailored to the individual user.
Meanwhile, two major revolutions are happening in science.
- The cost of generating personalized diagnostics is falling off a cliff. The cost of genome sequencing is dropping even faster than Moore’s law, and new kinds of diagnostics are entering the market, allowing many health signals to be detected earlier.
- The cost of manufacturing gene therapies customized for a single individual (n of 1) is also falling. We can now design and deliver personalized drugs via delivery vectors such as mRNA, and the FDA has shown a more open attitude toward letting patients try these procedures.
We believe these factors, combined, will bring a revolution in how care is delivered. Massive data and intelligence will help patients assess disease risk more accurately, so that treatment for the most serious diseases is no longer reserved for the few.
To make this happen, we believe there will be a large number of new startups covering every part of this ecosystem.
If you’re building products for a future of “intelligent personalized care,” we’d love to hear your thoughts too.
Company Brain
Author: Tom Blomfield
https://www.youtube.com/watch?v=IaWIazkWWog
The biggest factor holding companies back from AI automation is no longer the models themselves — models have become extremely powerful in a very short time. The real obstacle now is domain knowledge.
Every company has critical know-how scattered everywhere: some in employees’ heads, some buried in old email accounts, Slack threads, support tickets, and various databases. Companies still function because humans vaguely remember where this knowledge lives and how to use it.
But AI agents can’t work that way. If we want every company to run on AI automation, we need a new foundational capability: a company brain.
We need Garry’s G-Brain, but for every business in the world. A system that extracts knowledge from these fragmented sources, structures it, keeps it up to date, and turns it into a library of “processes and skills” that AI can directly act on.
This is not “enterprise search,” and it’s not “a chatbot over your documents.” It’s a living map of “how the company operates”: how refunds are handled, how pricing exceptions get approved, how engineers respond to incidents.
Then, AI systems can use this skills file to actually get work done in a safe and consistent way.
The company brain will become the missing layer between “raw company data” and “reliable AI automation.” I believe every company in the world will eventually need one.
If you’re building this, you should apply to YC.
Dynamic Software Interfaces
Author: Ankit Gupta
https://www.youtube.com/watch?v=_2LcG-3zCdY
Before AI, all users of a piece of software basically operated on the same interface. At most there was light customization, like a few different views, or theme and color options.
Even the “personalization” users think of with Netflix keeps the same layout for everyone — it just shows different images.
As a result, most software has a “one-size-fits-all” feel rather than being deeply tailored to the user.
For example: the way I use email is completely different from how most college students use email, yet all email clients look basically the same.
The exception is enterprise software: forward deployed engineers customize the software for each customer to make the experience fit better.
We believe coding agents are now strong enough that users can become their own forward deployed engineers, customizing the software they use far more aggressively.
I imagine a future where users design wildly different interfaces for different use cases — maybe my email client looks more like a task list, while a student’s looks more like an event calendar.
But these two interfaces would likely share some underlying primitives and design decisions, which software teams can build and deliver.
We believe future software companies will deliberately ship these shared primitives, assuming users will heavily reshape the final interface.
To realize this future, we must rethink the entire software delivery stack: how do developers make software accessible to users’ coding agents? Must they deliver source code instead of packaged binaries? Can agents only change front-end visual elements, or can they modify middleware on the fly to enable more interesting use cases?
If you’re a radical thinker who wants to define the future of software, we’d love to hear from you.
Inference Chips for Agent Workflows
Author: Diana Hu
https://www.youtube.com/watch?v=WQ8Q94mk22Q
Most AI chips are designed for a world where inference is “prompt in, response out.”
But that’s not how agents work. They loop: calling tools, branching, backtracking, maintaining context across dozens of steps. This is a completely different hardware problem.
Current GPUs only reach 30% to 40% of peak utilization on this kind of workload, because the load is bursty: bouncing back and forth between memory-bound model calls, I/O-bound tool usage, and CPU-bound orchestration.
This gap is where purpose-built silicon wins.
NVIDIA acquired Groq for $20 billion because it saw this trend. Google built the TPU v7 specifically for inference.
But no one is designing for the agent loop itself yet: fast context switching between models, native speculative decoding, memory designed for KV caches that must persist across the entire execution graph.
Groq’s real insight was not the chip, but the compiler that makes the chip work. We believe the same will be true for whoever does this next.
If you understand both chip architecture and how agents actually execute, this is a rare moment: experience on both sides matters equally.
If you’re building inference silicon for agentic AI, we’d love to hear from you.
SaaS Challengers
Author: Jared Friedman
https://www.youtube.com/watch?v=DhYJ1GENLoQ
Everyone is saying AI programming means the end of SaaS. Investors have wiped trillions of dollars off software companies’ market value.
That may be bad news for incumbents, but it’s good news for startups. If the giants are really this vulnerable, this should be the biggest startup opportunity in a decade.
So go be a challenger!
The SaaS model won because custom software was too expensive. A five-person team couldn’t possibly beat Salesforce.
But AI has cut the cost of producing software by 10–100x, and that changes everything.
The moats that once protected legacy SaaS — millions of lines of code accumulated over decades — are disappearing.
There are many ways to attack. The simplest: copy an existing product and sell it at one-tenth the price.
But you can go further: build an AI-native product from scratch — not by bolting a chatbot onto a 2010-era UI, but by fundamentally rethinking the workflow.
You can bundle ten SaaS point tools into one suite.
You can build an open source alternative to a product that costs $50,000 per seat, offer it for free, and monetize through services and hosting.
Most people will start with easy targets like project management tools. We encourage you to think bigger: go challenge the products that seem untouchable — chip design software, ERP, industrial control systems, supply chain management.
Those giant codebases of 10 million lines that no one has dared to touch for decades.
The last generation of great software companies was built by replacing on-premises software with the cloud. The next generation will be built by replacing legacy SaaS with AI-native software.
Software for Agents
Author: Aaron Epstein
https://www.youtube.com/watch?v=515J0QD9p6w
The next “trillion users” of the internet won’t be humans — they’ll be AI agents.
Now is the moment to “Make Something Agents Want.”
Agents are already browsing the web, doing research, placing orders, and managing legacy CRMs — but they’re working on software designed for “humans clicking buttons in a browser,” which is slow, inconsistent, and fragile.
Agents need a completely different foundation.
Instead of visual interfaces built around forms, buttons, and dashboards, they need machine-readable interfaces, such as APIs, MCPs, and CLIs.
Agents also need extremely thorough documentation, so they can discover, register, and immediately start using new tools programmatically, without any human involvement.
This means every major software category humans use today needs to be rebuilt for agents.
And the new agent-first software won’t come from incumbents “adding agent support on the side” — it will come from startups that explicitly build for agents as first-class citizens.
While everyone is building agents, the biggest opportunity may be building the software agents depend on.
So if you’re “Making Something Agents Want,” we’d love to hear from you.
Startups That Want to Sell to Huge Companies
Authors: Harshita Arora and Brad Flora
https://www.youtube.com/watch?v=Lrh1JVxqrKE
One of PG’s wisest pieces of advice has always been: startups should sell to other startups.
This has always been a “hack”: you quickly get smart, forward-looking users who help you refine your product into something that truly matters.
But it turns out there’s another class of companies with equally smart, forward-looking buyers that founders used to find hard to reach: massive enterprises.
We don’t mean “fairly big companies” — we mean the largest companies in the world, and these companies are often run by extremely smart, forward-looking people.
In the past, founders couldn’t do it: it was hard to reach the right people; hard to build a product “deep and complete enough” within a reasonable time for big companies to adopt quickly; and for big companies, the ROI of working with early-stage companies was too low and the risk too high.
AI has changed all three at once:
-
These companies — again, not just “big companies,” but “the largest companies in the world,” F100 scale — their executives are no longer hiding behind their computers. They are actively looking for teams that can solve critical problems with AI. In the past 3 years, we’ve seen YC companies land pilots and real contracts worth millions of dollars within their first year — even during their YC batch — for the first time. It’s no longer rare for a company’s first customer to be one of the largest companies in the world. The buyers are awake, and willing to talk.
-
For the first time, AI lets small teams deliver very thoughtful, detail-rich products to large organizations in “months instead of years.” The old joke about “lurking for 3 years to reach feature parity with the giants before launching” is dead. Now a 2–3 person team can build something a Fortune 100 company finds useful while the ink on the incorporation papers is barely dry.
-
The leaders of these companies also understand where a company should create internal value, where it should outsource, and what happens if it doesn’t adapt.
Again: these companies are run by genuinely smart people.
We think this is an excellent environment for founders, and we want to support more teams doing deals with the world’s most famous and important customers.
The AI Operating System for Companies
Author: Diana Hu
https://www.youtube.com/watch?v=K46ZNekw2Tk
The best AI-native companies we’ve seen have solved a problem many haven’t: they’ve made the entire company “queryable.”
Every meeting is recorded, every ticket is tracked, every customer interaction is captured — and all of it can be understood and learned from by an intelligence layer.
This turns a company from an open loop into a closed loop. In an open loop, you make a decision and maybe look at the results weeks later; in a closed loop, the system continuously monitors what happened, compares it against what should have happened, and adjusts.
I’ve seen teams that did this cut their sprint time in half and double their output.
The problem is, building this today requires painfully tedious integration work: stitching together Slack, Linear, GitHub, Notion, call recordings, and a dozen other tools with custom glue code.
There’s no product yet that connects all this context into a single intelligence layer — one that can reason across domains, alert engineering teams that they’re doing the wrong thing, or generate specs that agents can execute.
We think there’s a huge opportunity: build a “connective layer” that makes companies readable to AI by default.
Not another dashboard. A system that turns a company’s various artifacts into a self-improving closed loop.
Follow the WeChat official account Tech Backyard for more
