I’m increasingly convinced that the ability to build applications is being democratized in front of our eyes. Factors like creativity, data or taste – not necessarily technical skills or a massive budget – will become the deciding forces behind the next hit products.
And I’m not the only person who believes this. Replit founder and CEO Amjad Masad told me that their AI agent has made more than two million apps (!) in the past six months – all without requiring users to write a single line of code. 100,000 of these apps are hosted in production and being used by people, including for enterprise use cases like Zillow’s customer routing system. The company is now the single largest user of Anthropic models by tokens on Google Cloud.
Oh, and Replit’s subscriber base has been growing 45% monthly since the release of its AI agent. This is the fastest growth in the company’s history (Replit was founded in 2016).
I sat down with Amjad to unpack Replit’s (unhinged) growth journey and his unconventional learnings along the way.
Welcome to Growth Unhinged! Subscribe to join 68,000+ readers who get posts like this delivered straight to their inbox on Wednesdays at 6:50am Eastern Time.
Building my own app with Replit
Am I a software developer now? Certainly not, but I didn’t think I could understand Replit’s journey without seeing the product first-hand. I also wanted to try this whole vibecoding thing.
So I opened Replit’s website. In a matter of a few simple prompts, I had created a prototype SaaS pricing strategy advisor app – pulling from a website address along with advice given by yours truly in this newsletter.
The Replit AI agent proposed what to build including suggesting additional features I hadn’t considered like enhanced website analysis with machine learning and competitor pricing analysis.
I approved the plan and Agent (we’re on a first name basis 😉) was off to the races – quite literally. Replit shows progress in real-time and the coding was quickly writing itself on the right hand side of the tab. It’s all auditable, building trust and allowing technical users (aka not me!) to jump in and edit as they see fit.
When the initial prototype was built, Agent pinged me to test it. It specifically asked, “Can you see the Pricing Strategy Advisor interface with a URL input field? Does the page load correctly with the title and description?”
Not only could I see it, I was surprisingly pleased with the initial look and feel. It correctly scraped data based on the URL. And it generated a pricing strategy (“freemium pricing strategy with usage-based tiers”) does sound rather on brand.
The prototype took less than five minutes to create and test. It required zero coding experience. And the experience inspired me to go deeper – I’ve been thinking up both personal and professional apps that I would’ve never previously considered building myself.
The Replit origin story
Today, Replit positions itself as the easiest place to go from an idea to an app that you can deploy at scale. The product allows knowledge workers, aka anyone who sits in front of a computer, to take care of everything from setting up a dev environment to coding to hosting.
While the company has a big vision, the folks who get the most value today are still developer-adjacent – think designers, PMs, technical PMs, engineering managers, and executives. These teams use a large number of SaaS apps, yet still find it challenging to get exactly what they want so they complement those apps with low/no-code automation. Replit’s AI agent now gives these folks even more power – without the steep learning curve of even a low-code product.
Back in 2011 Amjad was a founding engineer at Codeacademy, which taught 50 million people how to learn to code (for free). After a run at Facebook working on React (a free and open-source front-end JavaScript library), Amjad founded Replit in 2016 with a focus on creating an online cloud development environment.
“It was way harder than expected,” Amjad told me. “It was very hard to raise money... We were lost in the desert for three years trying to bootstrap the company.”
Eventually, Replit got noticed by the likes of Paul Graham and Sam Altman. Despite being rejected by Y Combinator three times, Replit was finally invited to YC in 2018. The company’s ambition steadily grew and they raised a seed round from a16z. Replit hit another inflection point during Covid, but even still Amjad believed the company was far from realizing their vision.
By 2024, Amjad decided that the technology was there for an AI coding agent. He pivoted Replit’s roadmap to reflect that and launched early access in September 2024. Growth has exploded since.
A major driver: Replit’s agent-first user experience.
The UX decisions behind Replit’s breakout growth
One of the things that immediately caught my eye about Replit is that it was simple enough for an everyday person to use, but still powerful enough for fairly sophisticated teams.
Amjad explains that this UX decision was intentional. “We are big fans of progressive disclosure,” he explained. “This is the UX concept of being able to not overload new users with a lot of complexity while making the product so that advanced users can find ever more power.”
Most software products either expose users to “a shocking amount of complexity at first blush” (see: Adobe Photoshop) or they are “extremely simple, but constraining” (see Apple Notes, Canva).
Replit was inspired by tools that were powerful yet still approachable (see: Notion). Amjad believes that executing this well requires “a lot of taste” alongside making technological innovations.
At Replit, powerful-meets-approachable looks like having a fully natural language interface while always being two clicks away from opening the code editor or the operating system. Users can choose to peel open these additional layers one-by-one as they want to go deeper.
Amjad’s advice for others: being able to convey your value proposition on someone’s first visit to the website is critical. And you want users to be able to get to initial value as soon as possible.
He explained that he can see this in Replit’s data; the longer it takes for someone to get a “wow” effect, the less activation and the lower retention they’ll have. “We want you to get your first agent experience early on, ideally in minutes.”
Replit’s initial “wow” moment is when a user has created something, even if it’s a little contrived. Someone can click on the “habit tracker” button, for instance, and it’ll always create the same thing – sending the user through the exact same experience that Replit knows will be successful.
There is a higher bar for activation – this happens when the user types in their own prompt rather than using a canned prompt. That way the user is on their way to some success criteria where they can go from an idea to a working app.
Amjad emphasized the importance of exposing users to the right next action. “Even if people are successful in the first activity, it’s easy to get stuck and not know what to do. AI agents can suggest next steps, features for apps, whether to add a database, how to host the app, or if it’s time to invite a team member.”
Replit’s best (and worst) growth experiments
You might be surprised to learn that Replit has had breakout growth without a dedicated growth team. Amjad confided that Replit tried this, but it “had a negative impact on the volume of experiments” at Replit.
Replit has instead shifted to a decentralized approach and has put effort into democratizing access to the tools to be able to run experiments. (It’s worth noting that even core product improvements are A/B tested to ensure the team is making real progress.)
“There’s always a trade-off with specialization, which abdicates responsibility. Now we have data people who are embedded with teams and can consult on how to run experiments and to run the numbers to ensure statistical significance. And we increasingly have automated tools to show teams the metrics per experiment.”
The winning experiments have come from a mix of product improvements, pricing changes, messaging adjustments and UI simplifications:
Improving the core agent experience: Replit recently saw a 10% lift in app deployment after adding a reasoning model.
Pricing experiments: Replit tested entry price points of $20 per month (the control) vs. $25 per month vs. $30 per month. Amjad found that net-net $25 per month made the most money even after factoring in a slight dip in conversion. ($30 per month led to too much of a dip in conversion.)
Messaging adjustments: The team is continuously making small tweaks on the website to remove any jargon. “We used to call projects ‘repls’ and now call them ‘apps’. That was also A/B tested and people were more likely to activate and retain.”
UI improvements: Replit has regularly run experiments to reduce the number of clicks required to deploy an app. These often net out a 10% lift in completing the action with improvements flowing through to activation and retention.
Not everything has worked as expected. For instance, Replit has a number of integrations that they’ve trained the agent to be good at. They decided to expose these integrations under the prompt.
“Weird user behaviors started happening where they’d click on the integrations and not actually type in a prompt,” Amjad recalled. “This led to way fewer successful sessions. For now the agent will map your prompt against an integration that we think will work.”
Picking a hybrid pricing model with subscription and usage
Replit’s current pricing starts at $25 per month (monthly) or $15 per month (yearly). Each plan includes a certain amount of monthly credits with additional usage available on a pay-as-you-go basis. As I dug under the hood, credits can be used for AI agent checkpoints ($0.25 each) or edit requests with the AI assistant ($0.05 each) along with app deployment capabilities like compute, storage and data transfer.
I asked Amjad about the decision to pick this type of hybrid pricing model – particularly when there are so many flavors of AI pricing.
He attributed it to (a) margin considerations, (b) major differences in usage patterns across customers, (c) a desire to create pricing transparency for users, and (d) many of Replit’s products already being classic usage-based products (think: paying for compute).
“AI is one of those infrastructure pieces that’s very valuable where one user can use almost nothing and another user will be constantly chatting with ChatGPT and have negative margin,” Amjad reflected. “Many companies deal with this by adding ever more tiers – like ChatGPT adding a $200 plan. We thought value-based pricing is the right way to go and it creates more pricing transparency for users.”
As far as credits or tokens are concerned, it was important to Amjad to be transparent about what actions lead to spending credits. For Replit, he believes the most valuable action when working with an AI agent was a “checkpoint” – i.e. when an agent takes an action and it succeeds, it does a checkpoint or commit and then charges $0.25 per checkpoint.
Amjad was quick to emphasize that pricing, packaging and billing needs to be continuously looked at. “Most people think it’s a chore. But it’s an ongoing project and eventually companies will have a team that will continue to experiment and find a better user experience.”
Related: From selling access to selling work (and what it means for you)
Maintaining a lean team of generalists
Replit has raised more than $250 million in outside capital. The team? A mere 65 people. How it looks:
Engineering: 40 people
Design: 3-5 people
PM: 2 people
GTM: 5 people
The rest: HR, finance, legal, recruiting, etc.
GTM is a very new department at Replit. They have a head of sales and marketing. Reporting to the GTM leader is (a) one sales engineer, (b) one marketer, and (c) two BizOps people who are generalists who do a mix of growth, experiments, and talking to customers. In other words, the GTM team is exceptionally small for a business of Replit’s scale.
“We really like generalists at Replit,” Amjad reinforced. “They’re highly leveraged. And they know how to use Replit to automate some of their code.”
Looking at how GTM teams use Replit, sales engineering has been the most successful use case so far. When a sales engineer meets a customer, they can use Replit to quickly prototype an app. This allows folks to show highly personalized value to the customer with customized workflows.
The TL;DR:
Building AI agents as products is uncharted territory. Replit has had to reinvent the user experience, rethink traditional SaaS pricing and packaging, and re-conceptualize the best organizational structure for this new age.
As Amjad reflected, he told me: “For the first time since mobile you have a new interaction paradigm that’s fundamentally new. This is really exciting for product people.”
Are you vibecoding with tools like Replit? Drop a comment or a reply to share what you’ve built.
I rebuilt my company's website and deployed it on replit. It was an amazing experience, but also it's not doable for everyone. I definitely needed my analytical skills to debug.
Productsciencegroup.com
Great read particularly the GTM piece. So interesting. Believe that to make unique and deeper stuff you need to know code, but that will change i think. 1 billion unique customer apps. Sounds like a super support company is coming soon?