Y Combinator: Hardware, Agents, and Founders (Series #AI)
In The Lightcone episode “The State of Startups in 2026,” the YC team connects several changes through one idea: AI lets small teams take on more complex problems. Two things interest me here: what drives the economics of new manufacturing businesses, and why every software product would need its own agent. I would challenge that second proposition.
According to the figures the hosts present in the September 17, 2026 episode:
- hard tech rose from 8% to 20% of companies accepted into YC;
- robotics grew from roughly 1% to 6–7%;
- companies with a solo founder increased from roughly 5% to 18–19%.
These figures describe YC’s selection. They reflect both what founders propose and what investors prefer; I would not extrapolate the percentages to the entire market.
1️⃣ Hardware and demand for manufacturing
Here, hard tech includes robots, chips, energy, space, and defense systems. YC’s argument is that code generation lets a smaller team handle the software side of a complex physical product. That sounds reasonable, although the episode does not calculate savings across the full development cycle.
The Nox Metals example is interesting for a different reason. According to the hosts, new defense startups in the US need metal, and existing suppliers cannot keep up with their pace. Nox is developing automated metal processing and supply. Local manufacturing creates a market for the next link in the supply chain. Automation helps serve that demand. Attributing all this growth to cheaper development through AI would go too far.
2️⃣ A harness inside every product
YC then argues that systems holding business data should become environments where agents execute tasks: harnesses. Otherwise, an external agent will organize work with their data and may also take over the interaction with the user.
I do not believe this approach will succeed widely. Building a proprietary harness could consume a lot of money while users keep working in the Claude or Codex environment they already know. Through MCP and APIs, an external agent could potentially span more systems: a single task can easily involve a CRM, email, documents, and finance.
I would be more inclined to invest in product actions an agent can understand: check contract terms, approve a discount, place an order. With access controls, checks, and a way to investigate errors. A product’s own agent would still have to prove that working inside it delivers noticeably better results.
3️⃣ Experienced founders are getting attention again
The hosts highlight strong founders in their late 30s, 40s, and even 50s. Their explanation is that AI helps turn ideas into products faster, while experience helps identify what is worth building. They also suggest that managing engineers helps founders assign tasks to agents and assess their work.
I see a plausible mechanism here: knowing an industry and its problems becomes more useful when you can test a solution faster. But the episode offers observations and examples; it provides no success-rate statistics by age.
4️⃣ Starting alone has become easier
The increase from 5% to 18–19% refers to solo founders: people entering YC without a co-founder. The figure says nothing about employee headcount. The hosts themselves expect many successful solo founders to add co-founders later.
Their reasoning is that AI lets one person cover more tasks at the beginning. What interests me is the possibility of testing an idea and demand first, then assembling a team around work that is better understood. For someone with industry experience, that is a concrete expansion of what is possible: potentially fewer organizational obstacles before the first experiment.
The State of Startups in 2026 — Y Combinator
#AI #Robotics #Agents #Product #Management #Software #Engineering