Y Combinator's Spring 2026 batch had its Demo Day on June 16. Four days before it, I read through YC's directory one company at a time: 196 startups, 34 of them hiring anyone at all, the other 162 with no open roles. The morning after Demo Day I checked again, before any of the new raises had closed, and the count had not moved. I run a software company that sells engineering work, so that number is either my problem or my next market, and I wanted to know which.
Here is the plain version of what I found. The companies that usually tell you how software will be built next have built themselves with AI coding agents and very small teams, and they are not adding people to keep going. The most common team size in the batch is two. When one of these startups does post a job, it is almost always a founding engineer, and almost never anyone whose job is to hire other people. That pattern is deliberate, and it reaches well past this one batch.
A YC batch is worth reading because it runs a few years ahead of everyone else. The tools, the team shapes, and the unit economics that look extreme inside a batch tend to look ordinary across the rest of the industry about two years later. This cohort demos this week, and it is the first clean one assembled after coding agents stopped being a novelty and became the default way these founders write software. The shape of it matters more than any single pitch.
The number that started this
I did not get the hiring figure from a newsletter. I read it off YC's own directory, which exposes a hiring flag for each company, and then I read the same flag for three older batches on the same afternoon so I had something to compare against. Spring 2026 sat at 34 of 196, a little under 17 percent. Winter 2026, one batch older, was at 27 percent. Spring 2025 was at 38 percent. Summer 2024 was at 43 percent.
One honest caution before you read too much into that slope. I read all four batches on the same day, not at the same age, and the older ones have had more months and more raised money to switch hiring on. So this is a snapshot, not a controlled study. It still points one direction, and nothing else I dug up pointed the other way.
The team sizes say the same thing from another angle. Two people is the single most common size in the batch. Roughly five in six companies list four or fewer. The largest team in the entire cohort is nineteen. These are mostly not startups that raised a round and went quiet on recruiting; most of them never staffed up to begin with.
When they do hire, look at what they buy
The 34 exceptions are the part worth studying, because the kind of role they post is consistent. On Work at a Startup, filtered to this batch, the open engineering roles outnumber every other function put together, and more than half carry the word founding in the title. Founding engineer, founding robotics engineer, member of technical staff. These are equity-heavy, first-five-people seats, not capacity you rent to clear a backlog.
| Open roles in the batch | Count |
|---|---|
| Engineering | 9 |
| Marketing | 3 |
| Sales | 2 |
| Product | 2 |
| Operations | 1 |
| Design | 1 |
| Recruiting | 0 |
Those counts come off the live job board filtered to the batch, so treat them as the current sample rather than a full census. The shape is still hard to miss. Plenty of engineering, and not one recruiting role. A company that is not hiring recruiters is a company that does not plan to hire much of anything. When this batch brings someone on, it is not buying hands to do more of the same work faster. It is buying a co-founder it has not met yet.
A job posting is the most honest thing a company publishes. It is a bet it has already paid for.
What they are building instead
If they are not hiring people, it is worth asking what they are building with the agents instead. I took every company's one-line description, all 196 of them, and sorted each by a single question: whose work does this serve, the software and knowledge economy, or the real, physical one. The rubric is a judgment call, and I made one deliberate choice: software sold into a physical industry counts as real economy, so an ERP for manufacturers lands with manufacturing, not with software. Even under that conservative rule, the split is not close.
A slim majority, 106 companies, sell into the software and knowledge economy: tools for engineers, plus agents for recruiters, salespeople, marketers, support teams, analysts, and back-office staff. Read that list again, because it is the same list as a company's org chart. The single biggest thing this batch builds is software that does the desk jobs other companies used to hire for. The empty job board and the product are the same fact seen twice.
More than half the batch is building the thing that lets every other company stop hiring.
The real economy is not absent, which surprised me. 75 companies, 38 percent, build for physical and traditional sectors, and the weight sits in a few places: roughly two dozen in healthcare and biotech, around fifteen in manufacturing and heavy industry, and about ten each in defense and in robotics or hardware, with the rest scattered through logistics, construction, property, and government. These are the atoms bets, the ones you cannot ship from a laptop in an afternoon. Consumer apps, the thing an earlier YC era was built on, are down to thirteen.
You might expect those physical-world companies to hire more, since you cannot pour concrete or run a clinic with two people and a model subscription. Barely. The hiring rate is 19 percent among the real-economy companies and 17 among the software ones. Running lean is not a software habit in this batch. It is the house style.
The third map
People have been reading this batch two ways, and the two disagree. The pitch says agents; almost every company in the room described itself as one. The money says something else. By the accounts going around after Demo Day, the biggest checks skipped the agents and went to the atoms: a counter-drone company, a startup building return vehicles for things manufactured in space, a cancer scanner small enough to ride in the back of a truck. What the founders were building and what the investors paid for were two different things, and that gap is the argument everyone is having this week.
I read a third map, because I am not pitching the batch or investing in it. I am the one who sells the engineering, so I read it by payroll. Who is actually hiring, and for what.
Sorted by sector, software and the real economy looked close. Sorted by business model, they pull apart, and payroll lines up with the money, not the pitch. The motions that hire are the ones that touch the physical world or put their own name on the result: services companies at about 39 percent, deep-tech at 24, the handful of marketplaces at 25. The motions that barely hire are the pure-software ones, vertical SaaS at 9 percent, horizontal SaaS at 13, consumer apps at 7. Where the work is atoms or accountability, there are still people. Where it is software calling software, mostly there are not.
One detail in that chart is worth sitting with. The single biggest bet in the batch, 56 of the 196 companies, is the picks-and-shovels layer, tools and infrastructure sold to everyone else building agents. That is the layer the standard advice points at, sell shovels in a gold rush. It is also a low-hiring layer and, by plain arithmetic, the most crowded room in the building. One layer down from the agents is not open space. It is the most contested square foot in the batch.
They are not being cheap. They are spending it on tokens
The low hiring is not founders pinching pennies. It is where the money goes instead, and YC says this part out loud. On its own podcast, partners describe teaching founders a habit they call tokenmaxxing: treat spending on model tokens the way you treat rent, a cost you push as high as it will go, because an hour of agent time is cheaper than the work it stands in for. They talk, approvingly, about a founder dropping five hundred dollars on tokens in a single day.
The anchor story comes from YC's own president, who rebuilt his old blogging platform over about five days for roughly two hundred dollars in agent usage, against the millions of dollars and the team it took to build the first time, years earlier. Take the figures as the loose recollection they are. The ratio is the point, and the ratio is the whole premise of this batch: the thing that used to need a team now often needs a person and a credit card.
YC has wired the same belief into how it picks people. It now ships a tool that reads your Claude, Codex, and Cursor sessions and tells you, in plain marketing copy, that linking it improves your odds in its next program. Whatever you make of that, the signal is unambiguous. The people running the most-watched startup pipeline in the world are screening for fluency with coding agents at the front door, which is why a batch full of two-person teams reads as policy, not coincidence. One of its partners has put the destination bluntly: the next trillion users of the internet, he says, will not be people, they will be agents.
The work rhythm looks relentless from the outside. All-day agent sessions, founders complaining in public about hitting usage limits, a small tool economy growing up just to run several agents at once. I am not going to pretend to know their hours, because nobody in the batch actually wrote them down, and inventing the number would be worse than leaving it out. What is documented is the volume of machine work, not the human schedule behind it.
The part the demos skip
Now the inconvenient half, and it happens to be the half my business lives in. Agent code is not finished code. One founder in this batch, the technical cofounder of a healthcare startup, wrote up his own workflow and put a number on it: the code his non-technical cofounder ships with an agent lands somewhere around 60 to 80 percent of the way there, and then he spends an afternoon cleaning it up. He does that cleanup himself, in-house, and never once mentions hiring anyone to help.
That single sentence is the opportunity and the trap at the same time. The distance between 60 percent and production is real, and anyone who has shipped agent output knows the last stretch is the expensive one. But this batch is absorbing that stretch itself and paying nobody outside for it. A market does exist one tier down, where people without a technical cofounder pay to have AI-built apps rescued. Ulam Labs, an agency that has been around about a decade, now advertises a service line it calls "We Clean Up After Vibe Coding", and freelancers on Fiverr have sold the same repair for more than a year. The demand is real below the batch. Inside the batch, it is still being soaked up by the founders themselves.
The thing the agents cannot supply is judgment, and judgment is the last part of software to get cheap.
What this changes if you sell engineering
I am not going to land this on a reassuring note, because the data will not hold one up. If your offer is hands, more engineers writing more code, this batch is a preview of that demand thinning out, and not slowly. The agents really do write most of the code now, and the founders running them really are not hiring people to write more of it.
What does not get cheap is the work the Harbor cofounder described doing by hand: deciding what is correct, what is safe to ship, what falls over under real load, what the agent got confidently wrong. That is the layer we have spent the last year moving our own work toward, and it is the layer this batch is currently covering with one tired technical founder per company. Some of them will keep covering it alone. Some will hit the wall where an afternoon of cleanup is not enough, a security incident, a scaling failure, a regulated launch, and at that moment the question stops being how many engineers and becomes who actually understands this system.
That is a smaller market than selling hours, and a more crowded one, since the clean-up-the-AI-mess pitch already has takers. It is also the one that survives what this batch is showing us. Given the choice, I would rather sell judgment into a world that is shedding headcount than sell headcount into it.
If you are running an AI-built product and the afternoon-cleanup version of maintenance has stopped being enough, the production and correctness work is the part we do. You can see how we think about it on our AI engineering page.
