levelsio says YC's batch is harnesses and hardware only
Greg Isenberg looks at the same shift and says the work itself is now the product.
levelsio posted that YC's batch "literally only has harness startups or hardware startups", and drew the obvious conclusion. If Silicon Valley, the most early adopting place there is, has stopped starting software companies, that is the proof. His line was that software is mostly dead and hardware it is.
So software is mostly dead and hardware it is
He added the doubt that matters. It is debatable, he wrote, whether anyone will actually need a custom harness, or whether the frontier labs will just offer one generically. Garry Tan had already put the same shift more bluntly, saying you either die a system of record or live long enough to become a domain-specific harness.
The bull case, same facts
Greg Isenberg published the opposite read the same day. His definition of a harness is four things. It runs the model in a loop so it keeps working instead of answering once, gives it hands to read files and call tools, manages memory so hour three of a job knows what happened in hour one, and enforces rules about when it has to stop and ask a human.
a wrapper let you sell software ($800B market) but a harness lets you sell the work ($5T+ market)
His pitch is that a wrapper let you sell software, an $800B market, while a harness lets you sell the work, which he sizes at $5T+. Because the job knowledge lives in the harness rather than the model, you can swap GPT for Claude or an open model later and it keeps working. And because the harness knows when a job is finished, you can price per claim, per filing, per review or per closed month, which is a different competitive shape than SaaS.
Isenberg also argues the labs will not come for these. OpenAI, he wrote, is not going to learn how a freight claim gets denied in Rotterdam. He made the same point about OpenAI's Agents API, calling it an AWS moment where the hard engineering becomes a line item and the only thing left that is hard is knowing the job.
Where they agree
levelsio's practical advice this week pointed the same direction. Lots of money if you go IRL industries, he said, instead of selling tech to tech people who will not buy anything because they vibe code it themselves. Separately he argued normal users will never see a software layer at all, just a chat app they ask to do their bookkeeping or file their tax.
The business read. The doom call and the bull case both end up at the same instruction, which is to own one narrow painful workflow in an unglamorous industry, because the plumbing is now rentable and the domain knowledge is not.
Either you die a system of record or you live long enough to become a domain-specific harness