On November 24, 2025, I posted a short tweet pushing back on a specific promise that was going around at the time: that agentic coding tools give “non-developer folks” a lower entry barrier for “building a revenue generating app on their own.” Half an hour later, replying to a comment on it , I added the point that actually mattered more: vibe coding collapses the cost of building features, but it doesn’t collapse anything that lives outside the code: the data, the compliance, the trust, the distribution. Nine months later, that second point is the one I’d stand behind hardest, and the one the evidence has done the most to confirm.
What the tweet actually said
The framing was deliberately not the “AI lowers the barrier to entry, everyone gets to build their own app” story that was circulating at the time. My read was closer to the opposite:
“Although a lot of non-developer folks would like to believe this brings a lower entry barrier for building a revenue generating app on their own, this development eventually means higher production on a single developer, and ultimately it is challenging the SaaS model despite how shaky it already was.”
Two separate bets followed from that. On B2B: in-house developers could now produce more customized, more private, less costly versions of the SaaS tools their companies were already paying for, which pressures the vendor’s margin more than it helps a scrappy new SaaS competitor. On B2C: the same lowered barrier that lets a non-developer vibe code a revenue-generating app also lets someone else clone it a week later, so the endgame looks less like a wave of new winners and more like a pricing race, since a feature advantage stops being defensible the moment it’s cheap to copy.
Antigravity, the tool I used as shorthand for “agentic workflow,” had itself only shipped that same month. Google built it as a direct answer to Cursor. Naming a specific tool nine months ago was already a bet against how fast this category turns over, and that turned out to be the least durable part of the tweet.
The point that mattered more: moats live outside the code
The reply is the part I’d actually point people back to now:
“AI/ Vibe Coding/ Agentic Development collapses the cost of building features, but it doesn’t collapse the parts of software that live outside the code. I would suggest vibe coders to focus on structural moats when building towards commercialization: data ownership, regulatory compliance, liability and trust layers, multi-stakeholder workflows, enterprise procurement, and real distribution.”
The first tweet was about what happens to the SaaS vendors on the other side of this shift. The reply is about what a vibe coder actually needs if the goal is a durable, revenue-generating business rather than a working demo: not more features, since features are now cheap for everyone, but the six things a coding agent doesn’t generate as a side effect of writing good code: who owns the data, who’s liable when the software gets something wrong, how many parties have to sign off before a workflow can run in production, whether procurement will let the tool in the building at all, and whether anyone outside your own user base can find the thing.
The lineup changed, and not just around one tool
By mid-2026, comparison roundups routinely count seven serious agentic coding tools instead of two, and it’s worth being specific about how differently each one bet on what “vibe coding a revenue-generating app” should actually feel like, because none of them are simply a worse or better version of the others:
- Cursor stayed an editor first and added agents on top of it. Composer 2.5 leans on parallel task execution and broad model support, and it’s still the tool most engineering teams default to, largely on cost and polish rather than raw benchmark scores.
- Claude Code stayed terminal-native with no GUI at all, and kept winning the “hand it a genuinely hard, multi-file problem” comparisons, on the strength of long context and deeper reasoning rather than speed.
- OpenAI’s Codex, rebuilt on GPT-5.5 in April 2026 as OpenAI’s first fully retrained base model since GPT-4.5 , leans on cloud sandboxes with zero local setup and reportedly has around 4 million developers actively using it every week, the largest weekly user base of any tool in this list by a wide margin.
- AWS’s Kiro bet on the opposite instinct from vibe coding entirely: spec first, execute later. It makes you write requirements and a task list before it touches code, then its Autonomous Agent can take a ticket, run unattended, and hand back a finished pull request.
- GitHub Copilot stayed the cheapest and most ecosystem-embedded option, and added its own asynchronous “Coding Agents” mode so it could compete on the same terrain as the newer entrants without asking anyone to leave GitHub.
- Windsurf didn’t survive as itself. Cognition, the maker of Devin, acquired it in July 2025 and relaunched it in June 2026 as Devin Desktop, folding it into an “Agent Command Center” for managing a fleet of local and cloud agents from one view. It’s the clearest example in this list of a tool from November’s era getting absorbed rather than iterated.
Antigravity’s case deserves more than a version number. By its own public benchmarks, Antigravity trails Cursor and Claude Code on raw coding accuracy . It isn’t winning this category by being the best single agent at writing correct code. What it did with the 2.0 relaunch at I/O 2026 was change what it’s competing on: it split the agent manager out into its own desktop application, separate from any code editor, built around the idea that the developer supervises a project-wide fleet of agents rather than reviewing diffs in a file. That’s paired with a CLI, an SDK for custom agents, scheduled background tasks, and, by several accounts, the fastest raw output speed of any tool in the category. Whether “supervise a fleet instead of edit a file” turns out to be the right bet is a separate question, but it’s a genuine change in what the product is trying to be, not a marketing refresh, which is why it’s worth a paragraph on its own.
The frontier chat models underneath all of this moved through several generations in the same stretch too. Anthropic is now shipping the Claude 5 family. None of the specific names in my original tweet would be the ones I’d reach for today, and that turnover rate is itself the more durable point than any single tool’s feature list: the tooling layer is reshuffling roughly every quarter right now, which is exactly why the reply’s argument matters more than the first tweet’s. If the tools themselves can’t hold a lead for two quarters, nothing about which one you used is going to be your moat.
The B2B half showed up in earnings calls and stock charts
This is the part of the first tweet that stopped being an argument and became a data point, and it didn’t come from any single tool in the list above. It came from Anthropic pushing the same agentic pattern past code entirely. Claude Cowork, which launched January 12, 2026, extends agentic work into legal, sales, finance, marketing, and data tasks, and between that launch and February 23 the S&P Software & Services Index lost 25% of its value, more than $285 billion in market capitalization, per The Batch’s coverage . JPMorgan’s own software index fell 7% over the same window; LegalZoom dropped almost 20%, Thomson Reuters 16%. Wall Street started calling it the “SaaSpocalypse” before the month was out.
Underneath the headline number sits the exact mechanism the first tweet described. Retool’s 2026 Build vs. Buy report (Retool’s own survey of its customer base, so worth reading as a self-interested data point rather than an audited figure) found that 35% of its enterprise customers had already swapped out at least one SaaS subscription for a custom build, with most expecting to build more before the year is out. It doesn’t matter which of the seven tools above a given team used to do that. That’s the “in-house developers produce a more customized, more private, less costly version” line, showing up as a citable statistic instead of a hunch.
The B2C half is still the open bet
The consumer-side prediction (the pricing race breaking out once a revenue-generating app is cheap to build and cheap to clone) hasn’t gotten the same kind of confirmation. There’s no equivalent “$285 billion of B2C SaaS value evaporated” number to point to nine months in. What’s visible instead is the mechanism running one level down, in the tools themselves: seven different companies each shipped a competent coding agent within about a year of each other, landing within a few points of one another on adoption. Building a good coding agent is a genuinely hard engineering problem, and seven well-funded teams still converged on one that fast. A consumer app built by a single vibe coder is a much easier target to copy, so it should get cloned faster still. Whether that turns into an actual pricing race for consumer apps is still playing out, and, as the next section covers, the reply’s answer to that question is probably distribution, not price.
Which moats actually held up, nine months in
Take the reply’s list one at a time.
Data ownership and regulatory compliance turned out to be the most defensible pair, and the clearest confirmation is the same event covered above: on February 24, Anthropic didn’t route around Docusign, FactSet, Gmail, Intuit, and Salesforce, it partnered with them, because those platforms sit on data and audit trails a plugin can’t just re-derive by writing better code. The compliance side hardened further than either tweet anticipated: the EU AI Act, DORA, and the revised Product Liability Directive have pushed financial services, healthcare, and public-sector buyers to formalize AI third-party risk management this year, and ISO/IEC 42001 is increasingly showing up as a baseline requirement in enterprise AI procurement rather than a nice-to-have. None of that is something a solo developer generates by prompting an agent for a weekend.
Liability and trust layers moved from a line in a tweet to a documented gap in contract law. The same legal analysis argues the standard SaaS contract provisions covering data ownership, confidentiality, liability, and audit rights were written for static software and don’t hold up against continuously evolving models and layered sub-processor chains. Enterprise buyers evaluating agentic tools are now explicitly weighing safety governance, data handling, and vendor jurisdiction as separate line items before signing anything, according to ongoing tracking of the enterprise agentic AI landscape . That’s exactly the kind of scrutiny a vibe-coded feature can’t shortcut regardless of how good the underlying model is.
Enterprise procurement is the same Retool statistic from earlier, read differently: 35% of enterprises replacing a SaaS tool with a custom build is also a statement about which 65% didn’t, and procurement, not engineering capability, is usually the reason why. Getting past a buyer’s risk team is still a relationship and process problem an agent doesn’t solve by writing code faster.
Real distribution is where 2026 industry analysis has landed hardest on the reply’s side of the argument: one widely circulated founder-moat survey concluded flatly that distribution is now often a stronger moat than the AI itself , precisely because the cost of building a feature has collapsed for everyone at once, which makes who can actually reach users the scarce resource instead. That’s the same mechanism the B2C section above is still waiting to see resolved through pricing. Distribution may turn out to be the thing that decides it before pricing does.
Multi-stakeholder workflows is the one item on the list I can’t point to hard evidence for yet. It’s a real friction, production software that multiple departments or approval chains touch is much harder to displace than a single-user tool, but nine months in, I haven’t seen anyone measure it the way Retool measured build-vs-buy or the market measured the Cowork selloff. Worth flagging as unconfirmed rather than pretending it’s settled.
Five out of six holding up with citable evidence, nine months after being typed into a reply, is a better hit rate than most predictions get. The throughline is the same across all of them: none of these moats are things a coding agent, however capable, generates as a side effect of writing code. They get built the slow way regardless of how fast the code gets written, which is exactly what the reply argued before there was any data to back it.
Nine months on
The part I’d stand behind without changing a word is the core mechanism: a lower barrier to building software is not the same thing as a lower barrier to defending a business built on it. I expected that tension to play out over a few quarters on the B2B side; instead it showed up in about eight weeks, with a name, a $285 billion figure, and a follow-up correction attached to it by the same company that triggered it. But the reply is the one I’d tell a vibe coder to actually act on. If you’re building toward revenue and not just toward a demo, the code was never going to be what protected you, and nine months of earnings calls, contract law, and moat surveys have only made that argument more literal, not less.
This is a follow-up to a short tweet and its reply , both posted November 24, 2025. More of what I’m building at SATELYX .