On September 13, 2026, indie-maker-turned-commentator Pieter Levels — the person behind PhotoAI, InteriorAI, and Nomads.com — looked at a Y Combinator batch and didn't like what he saw. Replying to technologist David Galbraith's observation that "the only software in current YC cohort is domain specific harnesses," Levels went further: "software is mostly dead and hardware it is."
That line traveled fast, and it traveled with the kind of pushback that reveals more than the original claim. Within hours, other well-known builders — Marc Köhlbrugge of BetaList, Jon Yongfook of Bannerbear, and others — were arguing Levels had the framing wrong, or at least incomplete. The exchange is worth unpacking carefully, because underneath the "software is dead" headline is a genuinely useful question for anyone deciding what kind of AI company to build right now: what is a harness, why has it become the dominant YC startup shape in 2026, and does building one actually give you a moat?
TL;DR — the harness startup debate
| Question | Direct answer |
|---|---|
| What sparked this? | Levelsio claimed a YC batch was almost entirely harness or hardware startups, concluding "software is mostly dead." |
| What's a harness? | The orchestration/scaffolding layer around a frontier model — tool-calling, memory, context management, domain tuning. See the agent harness guide. |
| Is a harness "software"? | Yes, mechanically. Köhlbrugge's point: this is a new form factor, like desktop-to-SaaS, not software's death. |
| What's levelsio's actual claim? | Narrower than the headline: frontier labs may absorb vertical harnesses into their own generic offerings, eroding the startup moat. |
| Is "software vs hardware" the right axis? | Yongfook says no — much of the "hardware" in the batch may just be a physical wrapper around a harness. The real axis is moat durability. |
| What's the second threat to harness moats? | Open-source projects commoditizing specific harness patterns before any frontier lab even bothers. |
| What should founders do about it? | Ask what survives once (a) labs generically absorb the capability and (b) open source commoditizes the pattern — then build on what's left. |
What sparked this: a YC batch with no "plain" software
The immediate trigger was a post from David Galbraith, a technologist and VC, reacting to a video about a recent YC cohort: "the only software in current YC cohort is domain specific harnesses, but maybe even harnesses are general purpose. In which case, folks, VC backed software startups are largely dead." That thread reportedly drew close to 80,000 views before Levels quoted it with his own, blunter framing.
Levels' full post is worth reading in sequence, because it's an observation followed by a conclusion followed by a hedge:
"I just saw this @ycombinator's batch literally only has harness startups or hardware startups. So the proof is in the pudding. In the most early adopting tech pioneering place that's Silicon Valley new startups aren't even building software anymore. And it's debatable if anyone will actually need a custom harness or it won't just be generically offered by the AI frontier companies. So software is mostly dead and hardware it is."
Notice the structure: an observation about batch composition, a sweeping conclusion ("software is mostly dead"), and then — almost immediately — a hedge acknowledging the conclusion is "debatable." That hedge is where the real argument lives, and it's what the replies pulled on.
This isn't the first time explainx.ai has covered a "is software dying" flashpoint in 2026 — Chris Frantz's August viral thread argued teams were canceling SaaS to rebuild internally with coding agents, and the conclusion there was the same shape: software isn't dying, a specific layer of it is under price pressure. The levelsio/Galbraith exchange is a variant of that argument one level up the stack — not "will companies cancel their SaaS," but "is there a durable startup left to build once the model itself is this capable."
What "harness" actually means (and why it's become the default YC shape)
Before getting into who's right, it's worth being precise about the term everyone is arguing over, because "harness" gets used loosely. In 2026 AI-tooling vocabulary, a harness is the orchestration and scaffolding layer wrapped around a foundation model — the tool-calling loop, prompt structure, memory and context management, sandboxing, retry logic, and domain-specific tuning that turns a general-purpose LLM into a usable product for a specific task or vertical.
The model itself only predicts tokens. The harness is what wires those tokens into actions: parsing tool calls, executing them, feeding results back into context, and presenting output to a user. explainx.ai has covered this distinction in depth in the complete guide to agent harnesses — the short version is that Claude Code, Cursor, and Codex are all harnesses built around different (or sometimes the same) underlying models, and the same model can feel completely different depending on which harness runs it. A legal-research assistant, a customer-support bot, or a coding agent tuned for a specific codebase are all domain-specific harnesses in exactly this sense.
Why has this become the dominant startup shape YC is funding? Because the underlying models are now broadly available to everyone via API. If Claude, GPT, and Gemini are all a fetch call away, the model itself stops being a differentiator for a startup that isn't training its own frontier model. The near-term differentiated value — the thing an application-layer startup can actually own — currently lives in the orchestration and scaffolding wrapped around that model. That's exactly the argument YC's own September 2026 harness panel made explicit with a striking number: the same frozen model weights scored roughly 30% on ARC-AGI with a weak harness and about 95% with a strong one. Same model, wildly different outcome, based entirely on the scaffolding. If harness quality is where most of the near-term performance gap lives, it makes sense that harness quality is where most of the near-term startup opportunity lives too — which is a big part of why a YC batch would skew this way in the first place.
The actual disagreement: new category, or transitional layer?
Read past the "software is dead" headline and the real fight is narrower and more interesting: is a harness a genuinely new, durable software category — or a transitional layer that frontier labs will eventually absorb into their own generic product offerings?
The "still software, new form factor" position — Marc Köhlbrugge's reply cut straight to it: "Harnesses are still software. But just like software moved from Desktop to SaaS in 2000s, it's a new form factor harnessing (pun intended) the latest tech to make the best possible product." Klaas made the same point more bluntly: "aren't most domain-specific harnesses a saas/software" — pointing out a real tension in Levels' own framing. If YC is funding harness startups, and harnesses are software, then declaring "software is mostly dead" while sitting in a batch full of software companies is at least partly self-contradicting.
Levelsio's clarified position — pressed on this, Levels didn't retreat to "software is dead" as stated; he narrowed his claim considerably. To Köhlbrugge: "Yes just doubtful I think it won't be captured by the big AI companies either. Their models are big enough to do niches well. That's why it's completely different now than before." To Klaas: "Yes but I doubt they won't be part of general offering of AI companies anyway. Like Claude for medical etc." Once you read those replies, his actual thesis is narrower than the headline: it's not that harnesses aren't software, it's that vertical/domain-specific harnesses will likely get absorbed directly into frontier labs' own generic offerings over time — undermining the standalone-startup value proposition of building a custom harness as a moat.
Henno's supporting argument made the mechanism explicit: "It's gonna be pretty hard for any harness to outperform a frontier team over time, especially if frontier teams have infinite compute on models that aren't [siloed]." The logic is that a startup's custom harness is unlikely to durably outperform what a well-resourced frontier lab can eventually build into the base model or its own first-party product, given the lab's compute advantage and direct access to model internals a third party doesn't have.
Both sides have real precedent behind them. Köhlbrugge is right that desktop-to-SaaS was a form-factor shift that created an enormous new software industry rather than ending one — SaaS companies didn't get absorbed by Microsoft or Oracle just because those companies also shipped software. But levelsio's counter — clarified in his own reply — is that this cycle is "completely different" because the underlying models are themselves general-purpose and improving fast enough to absorb niche capability that used to require a dedicated harness. Whether that difference actually plays out the way he predicts is the open question; nobody in the thread, including Levels, claims certainty.
Yongfook's reframe: maybe it isn't software vs. hardware at all
The sharpest single contribution in the thread came from Jon Yongfook, maker of Bannerbear: "Haven't looked but I think even a lot of the hardware is essentially some meat space wrapper around a harness."
That's worth sitting with. If Levels' framing is "the batch is harness startups or hardware startups, and software is dead," Yongfook's reframe suggests the two categories might not even be separate — a "hardware" startup that's really a physical device shipping a thin interface to an underlying model-plus-harness stack isn't a fundamentally different bet than a pure-software harness startup. It just has an extra manufacturing and logistics layer bolted on.
This matters because it collapses the whole "software vs. hardware" framing Levels used to reach his conclusion. If a chunk of the "hardware" survivors in the batch are really harness startups wearing a physical shell, then the batch isn't split between two categories at all — it's overwhelmingly one category (harness-shaped bets) wearing two different form factors. The real axis worth asking about isn't software-vs-hardware. It's does this product have a moat independent of what the frontier labs will eventually ship natively — a question that applies identically whether the product is an app, a wearable, or a robot.
The faster threat: open-source commoditization
Even if you set aside whether frontier labs will eventually absorb a given harness pattern, there's a second, faster-moving threat to harness-startup moats: open source. Greg Wilk raised this directly in the thread — "Harness alone won't get them far. Especially with this being open source" — linking a GitHub repo as evidence that open-source alternatives were already commoditizing some of the exact functionality a harness startup might charge for.
This is a distinct mechanism from the "frontier labs absorb it" argument, and arguably a more immediate one. A frontier lab folding a capability into its generic offering typically happens on the lab's own roadmap timeline — months to years, and only for capabilities the lab decides are worth generalizing. Open-source commoditization can happen much faster: a well-funded or well-motivated open-source project can replicate a specific harness pattern — a particular tool-calling loop, a particular memory architecture, a particular domain-tuned prompt structure — in weeks, and once that pattern is public and free, a startup charging for the same thing loses pricing power immediately, independent of what any frontier lab decides to do.
explainx.ai's own coverage this year has tracked exactly this dynamic playing out inside the harness category specifically. YC's own QM — the multi-agent harness YC uses internally across accounting, legal, events, and engineering — was open-sourced under MIT in July 2026, immediately becoming a free alternative to any startup trying to sell a similar company-wide agent harness. The top open- and closed-source agent harnesses roundup documents how crowded and fast-moving this landscape already is on the open-source side alone. If a would-be harness startup's entire pitch is "we built a good orchestration layer," that pitch has to survive not one but two commoditization vectors — a frontier lab that might absorb it natively, and an open-source project that might give it away for free first.
What this means for what you build
If you're building an AI startup today, the levelsio/Galbraith exchange — stripped of its "software is dead" framing — surfaces a genuinely useful strategic question: what is my moat once (a) the frontier labs generically absorb this capability, and (b) open source commoditizes the specific harness pattern I've built?
A handful of categories hold up reasonably well against both pressures:
- Proprietary data that a frontier lab doesn't have access to and an open-source project can't replicate by definition — a dataset built from your users' actual workflows, not something scrapable or generatable.
- Genuine vertical distribution and relationships — regulatory relationships, existing enterprise contracts, domain trust built over years — that a generic "Claude for medical" offering from a lab doesn't inherit just by shipping a capable model.
- Physical-world integration that can't be replicated in software alone, which is arguably the more defensible reading of why "hardware" startups survived in the same batch Levels was reacting to — not because hardware is immune to harness commoditization (per Yongfook's reframe), but because the manufacturing, logistics, and physical-world reliability layered on top of the harness is a separate, harder-to-copy moat.
- Genuine UX and workflow innovation — a product that reshapes how a specific job actually gets done, not just a chat interface bolted onto a vertical prompt.
What's much less defensible, per this whole thread's own logic, is a startup whose entire pitch is "we built a good prompt-and-orchestration layer on top of GPT or Claude" with nothing else behind it. That's precisely the layer both of the thread's threats — lab absorption and open-source commoditization — are aimed squarely at. The same lesson shows up from a different angle in Zhenfeng Cao's "end of software engineering" paper, which argues that as agent systems get better at generating their own scaffolding at runtime, static code — including a lot of what a thin harness startup ships — becomes increasingly disposable tooling rather than a durable asset. And it's the same conclusion Diana Hu's YC playbook for AI-native companies reaches from the founder-advice side: token-maxing and thin middleware are not the moat; the operating discipline built around them is.
None of this means don't build a harness. Harnesses are real, valuable engineering — explainx.ai's own harness engineering deep dive shows benchmark gains coming from scaffolding improvements with the model held fixed, which is a genuine, teachable skill. It means treat the harness as the delivery mechanism for your actual moat, not as the moat itself. If the honest answer to "what happens to my business if Anthropic or OpenAI ships this natively next year, or someone open-sources it next month" is "we'd have nothing left," that's the gap to close before raising, not after.
Related reading
- What is an agent harness? Complete guide
- Is software dying or changing? Why teams rebuild internally
- YC's harness panel: self-improving agents, OpenJarvis, and QM
- YC open-sources QM: company-wide multi-agent harness
- NVIDIA Inception: Free Cloud Credits and VC Intros for AI Startups — a no-equity funding path worth layering on top of whichever go-to-market strategy a harness startup picks
- Garry Tan Wants an "American Distillation Regime" for AI Models — YC's CEO on why cheap open-weight models matter for exactly this category of startup
- Paul Graham's "Making Startups Powerful": The Full Framework Explained
- Top 10 open and closed-source agent harnesses (2026)
- Agent harness engineering: when the model stays fixed and the scaffolding wins
- The end of software engineering? Zhenfeng Cao's agentic paradigm paper
- Building AI-native companies in India: YC's blueprint meets bootstrap reality
- How to earn a billion dollars: Paul Graham's Oxford Union talk
- Agent Harness (dictionary)
Primary sources: Pieter Levels on X, September 13, 2026 · David Galbraith on X, September 13, 2026
This post reflects the public X/Twitter thread as it stood on September 13, 2026. Reply counts, view counts, and thread composition may have continued to evolve after publication; quoted replies are attributed to the handles that posted them and reflect the original phrasing.