Nine Months Later, Bubble’s AI Gap Is Compounding, Not Closing
Last June, I posted Why AI Threatens to End Bubble. The responses mostly fell into two camps: “this gets posted every week” and “AI can’t build anything serious.”
I want to revisit, but I want to start somewhere different this time.
The Real Question
Learning Bubble is hard. It took me months to get decent and over a year to feel confident. The editor, the workflow system, the data model, privacy rules, responsive design, API connector, workload unit optimization. That’s a serious investment, and anyone who’s been through it knows what I mean.
I also want to be clear: Bubble built my foundation. The way I think about data structure, application logic, user flows, and product design all came from building on Bubble. I wouldn’t be where I am today without it.
These foundational skills, thinking about databases, understanding how logic flows through an application, knowing how to scope a feature, these aren’t Bubble skills, they’re software skills. Every experienced Bubble developer already has the foundation needed to direct an AI agent effectively. You’ve been training for this without knowing it.
But there’s a question I keep coming back to: if someone is starting today, or deciding where to invest their next year of learning, should that time go into Bubble?
Two Curves
There’s a learning curve either way. Bubble has one. Agentic coding tools have one too. You still have to understand how databases work, how to think about deployment, how to structure a project, how to articulate what you want clearly. None of that is free.
The difference is what happens to those skills over time.
Bubble skills compound inside Bubble. You get faster in the editor, you learn the patterns, you build reusable components. But that knowledge is locked to the platform. It doesn’t transfer, and it doesn’t get more powerful unless Bubble ships new capabilities.
Skills outside Bubble, understanding tools like Supabase, knowing how to deploy to Railway or Vercel, learning how to direct an AI agent, live in an environment that is compounding exponentially on two axes simultaneously.
First, the tools are getting dramatically easier to use. Nine months ago, deploying an app or managing a database outside Bubble required more from you. Today, the friction has dropped significantly and continues to drop every month.
Second, and this is the part that really matters, the capability ceiling is rising dramatically. It’s not just that these tools are simpler, rather what you can build with them is expanding at a pace that has no historical precedent. Every model upgrade, every new agent feature, every infrastructure improvement immediately translates into more capability for anyone in this ecosystem. Your existing skills become more powerful without you doing anything additional.
That combination, easier to learn AND exponentially more capable, is what makes this different from any previous platform debate. This isn’t “Bubble vs. WordPress” or “Bubble vs. Webflow.” Those were lateral moves. This is a fundamentally different growth curve.
And it’s not a coincidence. Hundreds of billions of dollars are flowing into AI infrastructure, data centers, model training, and agentic tooling. The entire industry, OpenAI, Anthropic, Google, hundreds of startups, is tackling variations of the same problem: making it easier to build software with AI. Bubble is one company with finite resources trying to integrate AI into a proprietary system. That’s not a fair fight, and it doesn’t need to be anyone’s fault for it to matter.
What Changed in Nine Months
To make this concrete, here’s a side-by-side of what’s happened the past nine months:
Outside Bubble
Claude Code now runs multi-agent workflows where parallel agents tackle different parts of a task simultaneously, operating directly on your codebase. Boris Cherny, Head of Claude Code, stated it generated roughly 80% of its own code. A Google engineer publicly acknowledged that Claude reproduced a year of architectural work in one hour. Microsoft has widely adopted Claude Code internally. It hit an estimated $2.5 billion in annualized revenue by early 2026, nearly doubling since the start of the year. Anthropic’s overall revenue reached a $14 billion run rate, up from $1 billion at end of 2024.
OpenAI Codex shipped four major model upgrades in under a year. The latest is what OpenAI describes as “the first model instrumental in creating itself.” The Codex App lets you manage multiple parallel agents on the same repo, and ships with Automations that handle issue triage, CI/CD, and monitoring unprompted. It surpassed 1.5 million weekly active users by February 2026. Cursor, the AI code editor, hit $2 billion in annualized revenue and is seeking a $50 billion valuation.
Replit shipped three major Agent versions in about a year, each delivering 2-3x speed improvements. Agent 4 runs parallel agents across auth, database, backend, and frontend simultaneously, tests its own code, works 200+ minutes autonomously, and builds other agents. Replit went from $10M to $100M ARR in 9 months.
This all largely made possible by drastic decreases in the cost of intelligence. Getting GPT-4 level intelligence cost about $20 per million tokens in late 2022. Today it’s about $0.40. AI of equivalent quality gets roughly 10x cheaper every year.
Inside Bubble
Bubble’s AI Agent launched in beta and is currently only available in apps built using Bubble’s AI app generator. It can create and modify UI, generate dynamic expressions, and work with data types. Workflow generation is being added.
Bubble’s agent still can’t handle conditionals, doesn’t support plugins or API-driven data sources, can’t delete or restructure data types, and can’t access logs or the issue checker. Native mobile editing is guidance-only. These aren’t edge cases. They’re core parts of building a production app.
Bubble’s app generator can scaffold an app in 5-7 minutes, but there’s no conversational iteration after that. As No Code MBA’s 2026 review put it: “Bubble generates the first version, then you’re on your own.” The agent is still rolling out to existing apps, with paying apps months away. MCP support is “on our radar, but not on the roadmap.” Workflow branching won’t arrive until H2 2026 at the earliest.
Emmanuel said it directly in his March 2026 AMA: “Transparently, our AI offering is not yet at par with some of the AI-first companies in our space.”
That’s an honest statement. But when you put the two timelines side by side, the gap didn’t narrow over the past nine months. It widened.
Why the Gap Compounds
This isn’t just about execution speed. It’s structural.
Bubble needs to individually engineer every new capability into its agent because the underlying system is proprietary. Conditionals, plugin support, backend workflows, each has to be built as a feature. Code-first agents just write code. Any new framework or pattern is immediately available. No engineering gate required.
Emmanuel’s counterargument: “Code gives you the illusion of control.” I’d flip that. Bubble gives you the illusion of unlimited capability within a controlled environment. You feel in control because you can see everything visually, but you’re constrained to what Bubble has built support for. The moment you need animations, real-time features, a specific API pattern, or custom server-side logic, you hit a wall.
With agentic tools, you spin up a local server and see your app running live as you make changes. You watch the agent edit in real time, test the result immediately in your browser, and tell it to undo or adjust anything on the spot. It’s visual, immediate, and interactive. And there’s no ceiling. The constraint isn’t the platform’s feature set; it’s your ability to describe what you want.
My Experience
A caveat: I may not be be the median Bubble user. I’m comfortable with architecture decisions, database design, deployment, and security. I don’t write code, but I’m willing to learn how systems fit together. That said, I think anyone who’s gotten good at Bubble, whether intermediate or expert, already has the foundational skills to do what I’m doing. If you can design a Bubble database, you can work with Supabase. If you can think through Bubble workflows, you can direct an agent through application logic. The jump is smaller than you’d expect.
I used Bubble for everything. Multiple production apps, years of investment, real revenue. When Replit Agent and Claude Code became capable enough, I started building new projects on them. I’ve now migrated roughly 90% of my apps off Bubble.
Here’s a concrete example from yesterday. I had a Bubble app I built about two years ago. It’s a fairly simple app, but it still took me a couple weeks to build originally. For the past six months, I’ve had a list of changes I wanted to make, optimizations, UI improvements, new features. I kept putting it off because iterating in Bubble’s editor is slow enough that it never felt worth the time.
Yesterday I decided to migrate it. I spun up a Railway server, created a Supabase database, and gave Claude Code the credentials for both along with my Bubble app export (.bubble file). I told it to analyze the export and create a plan to rebuild the app using Supabase as the database and Railway for deployment. It generated a full refactoring plan. I gave it my Bubble data API token, it exported all the data, and I ran the plan.
Within an hour I had a nearly identical copy of my app running locally. By the end of the day, it was fully deployed, the database was migrated, and I had made every change I’d been putting off for six months. One day versus six months of procrastination, not because the changes were hard, but because the friction of doing them in Bubble made them not worth starting.
That’s not an isolated case. The speed difference across everything I build is 5-10x or more. And beyond speed, there are hard limitations that simply don’t exist outside Bubble: animations, advanced UI, complex API integrations, MCP connections, custom server-side logic, real-time features. If there’s a library that does what I need, I use it.
I still don’t write a single line of code. But I understand systems well enough to direct an agent, and that skill set is getting more powerful every month while Bubble-specific expertise stays roughly the same.
Where This Is Headed
I’m not asking “will Bubble die.” But I think anyone invested in the platform should be watching the trajectory honestly and closely.
To be fair, I underestimated Bubble’s ability to retain its existing user base through switching costs. Existing production apps aren’t going anywhere soon. And the value of managed infrastructure for users who genuinely don’t want to think about ops is real. I was too dismissive of that in my original post.
But if the current trajectory holds, which every single thing is pointing to it doing so, fewer new builders will choose Bubble as their starting point and more and more users will start migrating off.
As code-native tools get easier and more capable simultaneously, the case for learning a proprietary visual system weakens. And that creates a cycle that’s hard to reverse: fewer new users means less revenue, less revenue means less investment in innovation, less innovation means the gap widens further, which drives away more users. Bubble doesn’t need to make a catastrophic mistake for this to happen. It just needs to keep improving at its current pace while the market around it accelerates.
Bubble is not failing. Bubble is improving. But it is improving on much a slower curve than the ecosystem around it, and that matters because AI shifts competition from “who has the best visual builder” to “who compounds model progress into real building capability fastest.”
For the Bubble team and community: what is Bubble’s path to closing this compounding gap before that cycle takes hold? Because from where I’m standing, it’s already turning.
That question feels much more urgent in March 2026 than it did in June 2025.
I would love to hear yalls thoughts
to fix such huge and deep problems with a quick one second push of my foot pedal… 
