← Gwen Working Papers

What Y Combinator Wants Built in 2026 — and Where Software That *Does* the Work Fits

August 1, 2026

Twice a year — now roughly every batch — Y Combinator publishes a Request for Startups (RFS): a public list of ideas its partners and founders would like to see built. YC is careful to frame it as an invitation, not a filter. "These represent just a fraction of what we fund," the page notes, "if one excites you, take it as extra validation to dive in, but you don't need to work on these ideas to apply." Read that way, the RFS is less a mandate than a mirror: it shows what the people who fund a large share of early-stage software think is newly buildable, and newly worth funding, right now.

The current list is the RFS, published on ycombinator.com/rfs. Its framing thesis is stated up front: YC describes a wave of startups "rebuilding the systems that power the real world, from education and healthcare to defense, finance, infrastructure, and work itself," and adds a first for the tradition — one request authored by a sitting U.S. Secretary of the Army. That detail signals how far the frame has shifted from consumer apps toward institutions, hardware, and the machinery of states and industries.

What YC is asking the world to build

As of this writing, the Fall 2026 page lists thirteen requests, each attributed to a named author:

Why the list looks the way it does

Three currents run through the whole page.

First, None of these requests argues that models are useful; they assume it and ask what becomes buildable next. Schaad's consumer request makes the economics explicit: capability is "good enough" to treat an agent like a person, and per-user inference cost — he cites figures on the order of ~$1,000/month in tokens today — is falling roughly 10x a year. That curve, more than any single breakthrough, pulls "someday" into "now."

Second, Defense hardware, offshore compute, home robotics for elders, sensors on balloons and pipelines, operating systems that route a job between an agent, a robot, and a person — the center of gravity is physical, regulated, and institutional. This is a deliberate turn away from the pure-software comfort zone, and it explains the recurring emphasis on proprietary, end-to-end data: as the physical-world OS request puts it, whoever records the work as it actually happens holds data the frontier labs and incumbents will not have.

Third, "Proving You're Human," AI-native compliance, and self-maintaining APIs are all, at root, about keeping systems honest and in sync as agents proliferate. When software starts acting on its own, the hard problems become verification, permissioning, and change management — the connective tissue rather than the raw capability.

Where software that the work fits the pattern

Read the list from a distance and a shape emerges: most of these requests reward systems that , not systems that . The Primer teaches, it doesn't recommend a curriculum. Multiplayer AI is valuable precisely because the agent hours- or days-long work a team steers in real time. The physical-world OS earns its keep by dispatching and completing jobs. Compliance infrastructure wins when it generates the report and keeps the audit trail, not when it explains what a rule means. Self-maintaining APIs are explicit about it — the request's whole thesis is that providers "shouldn't just announce changes; they should apply them."

That distinction — advising versus doing — is the through-line, and it is why an AI workspace built to deliver outcomes maps naturally onto several of these requests. Where a founder describes a result and the system builds the app, produces the media, runs the outreach, and does the operational follow-through, the "last mile" that most of these RFS entries are really asking for is already the product. Multiplayer AI, the cloud for small software, AI-native compliance, and the operating systems for real-world work all describe pieces of that same pattern from different angles: turn intent into completed work, shared across a team, with a durable record of what was done. Gwen — an AI workspace that builds and hosts web apps, generates images and video, runs email, CRM, and social, and carries out research and operations — sits inside exactly that gap between advice and execution.

That is observation, not a claim of coverage. Several requests — offshore compute, defense hardware, physical sensor networks — are firmly in the domain of atoms, and no software workspace answers them. The honest read is that "AI that does the work" is one strong axis of the 2026 list, not the whole of it.

A note on what's confirmed versus reported

Everything above about the requests themselves — the titles, authors, framing thesis, and the count of thirteen entries under the Fall 2026 tab — is drawn directly from the live ycombinator.com/rfs page and is confirmed against that primary source. Broader characterizations of the 2026 cycle that circulate in tech press (batch sizes, acceptance dynamics, or thematic commentary) are rather than confirmed here, and are not relied on for the specifics above. Where this article makes a claim of fact, it is the RFS page's own language.

What comes next in this series

This piece is the primer. Each of the thirteen articles that follows takes one request and goes deep: what YC is actually asking for and why now, what the hard technical and go-to-market problems are, who is already circling it, and — stated plainly and without hype — where Gwen does and doesn't stand on that specific request. The aim across the series is the same as here: credible analysis first, honest about what an AI that does the work can deliver today and where the request runs past it.

Try Gwen - the AI that does the work