YES-CODE is here. So where's the SaaS moat in 2026?

I was up late scrolling X when Guillermo Rauch, the CEO of Vercel, dropped a short thread. No hedging. No build-up. Just a one-word verdict.

YES-CODE.
An entire category of software, “no-code”, was built under the presumption that code is expensive, difficult, and scarce.
Coding agents have forever changed the equation. Code is now cheap, easy, and abundant. [1]

Within 24 hours, the same conclusion came back from three people who don’t usually agree on anything.

Shawn Wang (@swyx) — independent engineer, runs Latent Space — posted a screen recording of Codex finishing a non-trivial task on the first try. His reaction: “codex is agi man, oneshotted this, no notes.” [2] Engineers voting with their keyboards.

Peter Yang (@petergyang, product at Roblox, 140K-subscriber newsletter writer) brought it back to the business side [3]:

People are saying SaaS is not dead.
I think larger enterprise SaaS that can do multiple jobs are probably fine (e.g., Figma).
But if you’re building a simple SaaS for a narrow use case, I think it’s harder to monetize now because AI skills can often solve the same problem…

And then there was Aaron Levie, the CEO of Box, who didn’t pick a side. He reframed the whole question [4]:

As token budgets take on a larger part of operating expenses over time, model routing is the inevitable conclusion. This is also one of the biggest areas of differentiation for the applied AI layer over time.

A platform CEO, an enterprise CEO, a product lead, and an independent engineer. Four people, four seats, the same punchline. That stopped me.

My honest first reaction

I’ll be straight with you. When I first read YES-CODE, I wasn’t excited. I was uneasy.

I’ve already written about Claude Code’s /workflows and the agent middleware layer on this blog. Both pieces assumed the same thing: if AI makes code cheap, the moat migrates up — into agent orchestration, context engineering, model routing. One layer above “the code.”

What @petergyang said that actually slowed me down was a line buried inside a longer post [5]:

They’re so lit up by being able to build anything that they don’t do anything else. Then they launch and there are no users.

In plain English: builders are so high on “I can build anything” that they forget to check whether anyone wants it.

This isn’t a critique of the technology. It’s an honest autopsy of “shipping ≠ selling.” When AI makes shipping near-free, “why you, why now” jumps from the back of the room to the front.

So where is the moat, actually

I lined up the 24 hours of takes by role. Here is what each person is actually arguing for.

Notice what’s missing. Not a single person said “data” or “user count.”

Data moats are getting steamrolled. An agent can reconstruct in 24 hours what a SaaS spent three years accumulating. Switching costs are getting ground down by tools like Codex and Claude Code that fit wherever the work is.

What is left? Reading the full day of takes, here is my own ranking, weakest moat to strongest.

1. Workflow entrenchment (@petergyang’s lane)

If you are Figma, Linear, or Notion — products that knit several jobs together for a team — you are probably fine. You are not selling a feature, you are selling an organizational habit. Figma’s value was never drawing shapes. It was closing the design–dev–review loop.

Sounds obvious. But it gives you a sharp negative filter:

If your SaaS does one job, and that job can be replaced by a single prompt plus an MCP server, you should be panicking right now.

2. Model routing and token economics (@levie’s lane)

A lot of people retweeted @levie as a one-liner. The deeper read is different. The future question is not “which model is smartest.” It is “which model, how much token, in what order, for this specific workflow.”

That is also why GitHub Trending is being dominated right now by headroom, codegraph, and agentmemory. Same problem, three angles: let the agent do more with fewer tokens. Whoever turns model routing into a product surface — not a research paper — wins the OPEX moat.

This is sold as a technical story, but it is really a commercial one: if you can show a customer that your stack spends fewer tokens per task than what they could assemble themselves, that is real money.

3. Channel and scene ownership (the one most people miss)

Right after YES-CODE, @rauchg posted a separate endorsement for Conductor, an “ADE” (agent development environment):

Conductor has the distinct edge of being an IDE born for coding agents. An ADE if you will.
Agents will take remote dev mainstream.

Read the words carefully. He is not saying “we are building a new IDE.” He is saying agents will take remote dev mainstream. His bet is that local development is no longer the default; remote is.

Vercel’s real moat was never Next.js. It is the position it holds in the chain “an agent wrote code → the code ships → the world can hit it”. That is not a technical moat. It is a scene moat.

For SaaS founders, the takeaway is uncomfortable and useful:

Stop competing on features. Compete on being the first call the agent makes.

4. Taste and orchestration craft (@swyx’s lane — but don’t mythologize it)

@swyx’s “codex is agi man” reads like hype. It isn’t, really. It says: an indie engineer plus a competent agent can now deliver what used to take a team.

That opens a door for solo founders. It also closes one. The door opens because more things become buildable. The door closes because more people can build them, so differentiation has to come from “what the thing feels like to use,” not “the fact that it exists.”

After years of writing code and shipping products, one of the few lessons I trust is: in the age of AI, taste is the scarce resource. Spacing, naming, restraint, the way “polish” is preserved all the way through a flow — none of that comes from a single prompt. It comes from someone’s long-accumulated judgment.

Three things I would actually do, if I were a SaaS founder

Reading the 24-hour window end to end, here is what I would do this quarter.

One. Stop adding features. Start asking whether each feature has already been replaced by a prompt. That is literally the only claim @petergyang is making. If 80% of your core value can be rebuilt as a Claude Skill in 2026, every GTM dollar you spend is building a sandcastle.

Two. Stop selling “we have X years of customer data.” Data was the 2010s moat. In 2026, agents compete on judgment and routing, not on “we have 100M rows.”

Three. Take model routing seriously, whether or not you build it yourself. This means your product architecture should let you swap models per task. If your SaaS is hard-wired to a single provider’s API, the next price cut from Anthropic, DeepSeek, or Mistral eats your margin before you can react.

The questions I am leaving open

Three questions I haven’t answered for myself. Writing them down so I have to look at them again.

1. Is the thing I’m building already on @petergyang’s “narrow use case SaaS” list?
2. Will my customers, twelve months from now, feel like “I just had Codex do this for me”?
3. If I repriced tomorrow, how much of my invoice is “feature fee” and how much is “judgment / routing / orchestration brain fee”?

If you can answer those honestly, you don’t need anyone to tell you whether the moat holds.

References

[1] Guillermo Rauch (@rauchg), original post: https://x.com/rauchg/status/2061934154732974376
[2] Shawn Wang (@swyx), original post: https://x.com/swyx/status/2062062585391014245
[3] Peter Yang (@petergyang), original post: https://x.com/petergyang/status/2061846283263103274
[4] Aaron Levie (@levie), original post: https://x.com/levie/status/2061974298760495132
[5] Peter Yang (@petergyang), follow-up post: https://x.com/petergyang/status/2062018242789670929
[6] Guillermo Rauch (@rauchg) on Conductor: https://x.com/rauchg/status/2061809689973944724