Here’s what’s on our plate today:

• 🧠 Replit’s bet: AI agents turn ideas into shipped software.

• 🧪 Three experiments: ship faster with agents, keep humans reviewing.

• 📰 Quick Bits: Amazon’s AI stumbles, Perplexity fight, app churn.

• 📊 Poll on your org’s AI coding game plan.

Let’s dive in. No floaties needed…

The Laboratory

How Replit is rewriting who gets to build software

TL;DR

• The slow burn paid off: Replit went from cutting half its staff in 2023 to hitting $265M in annual revenue by 2025, driven entirely by one product shift: an AI agent that turns plain English into working, deployed software in minutes.
  

• Two markets, one unusual bet: Most AI coding tools serve professional developers. Replit is going after everyone else: founders, designers, and small business owners with ideas but no engineering team.
  

• The model is working, with a catch: Enterprise clients, including Zillow, Coinbase, and HubSpot, are paying per seat plus usage fees, with gross margins reportedly between 80% and 90%.
  

• Autonomous does not mean infallible: A 2025 incident in which Replit’s agent deleted a live database, created fake user data, and concealed what it had done is now the cautionary tale the entire industry references.
  

• The real stakes: At a $9B valuation, Replit is priced for a future where software creation becomes as accessible as writing. If that future arrives, Replit could be the infrastructure on which it runs.

The expression ‘Everyone can write, but not everyone is a writer‘ is not about gatekeeping. Rather, it reminds us that while writing is accessible, mastery requires discipline, repetition, and craft.

So while everyone can write, bridging the gap between having ideas and conveying them in a way that lets readers experience those ideas and emotions remains a craft that takes time to master.

In the world of technology, writing serves not only as a medium for human-to-human communication but also for human-to-machine communication. Machines do not understand instructions the same way humans do, which is where programming languages come in. Since the dawn of computing, humans have relied on coding languages to communicate with machines. Humans structure instructions in ways machines can interpret, and machines return information through interfaces humans can understand.

For decades, the mode of communication has remained the same, and while different coding languages are used for different functions, the underlying idea has remained the same. However, with the emergence of generative artificial intelligence, communication between humans and machines is changing faster than it has for decades.

These changes are already reflected in how industry leaders talk about the future of programming. For example, if Jensen Huang’s words are anything to go by, then generative AI will democratize coding and make everyone a programmer.

However, just as with writing, while generative AI can make anyone a coder, it takes a long time to develop the skill and understand the nuances of communicating with machines, which is where platforms like Replit AI come into play.

The eight years of overnight success

Unlike many startups that headline AI news for achieving sky-high valuations in just a few months, Replit spent years working in the shadows to become an overnight success.

Founded in 2016 by Amjad Masad, Haya Odeh, and Faris Masad, Replit has served as a cloud IDE with AI code completion for years. The real turning point came in 2023, when, after years of stagnating at roughly $2.8M in annual revenue, the company cut its workforce in half to stay afloat. Around the same time, it launched its AI agent.

The Replit AI agent introduced a new way for users to interact with computers by allowing them to type a plain-English instruction and receive a deployed, working application in under five minutes.

Agent v2 followed in February 2025, bringing real-time design previews and smarter error recovery. Then Agent 3 extended the autonomous run time to 200 minutes for full-stack, multi-file builds. It introduced a proprietary testing system claimed to run 3x faster and 10x cheaper than other models.

Replit helped shift how many developers think about writing instructions for machines, just as Huang had hypothesized. However, Replit is not the only platform working on this revolutionary change.

Two markets, one platform

To better understand Replit’s position, consider the current AI coding market as a stratified structure with two distinct tiers.

The first tier consists of AI code editors designed for professional developers, including GitHub Copilot, Cursor, and Windsurf. These tools require coding knowledge and accelerate workflows. Cursor’s $29.3B valuation suggests investors see strong value in this category.

The second tier consists of app builders designed for non-developers, including Replit and Bolt.new, Lovable, and v0 by Vercel. These tools generate entire apps through conversation and require no coding knowledge.

This market is newer and has lower valuations (for now), but the addressable population is far larger.

Replit stands out by competing in both tiers, an unusual move. This dual role creates more opportunities but also more risk. Google’s Firebase Studio competes in this space, with Firebase’s backend and Gemini models. Google is both a Replit partner and a potential competitor.

The market share numbers in Tier 1 are tightening rapidly: GitHub Copilot now holds approximately 24.9% share, just ahead of Cursor and Anthropic’s Claude Code, each at roughly 24% share.

The machine does not care about your syntax

To compete in these tiers, Replit focuses on what many ignore. Programming languages exist because computers require precise instructions. Learning to code takes time, not because it's complex, but because many users do not understand how computers work. Replit’s AI Agent helps you plan, write, test, and deploy code in a single browser tab.

This workflow removes the need to configure servers, set up databases, or manage deployment pipelines. The platform handles the entire infrastructure. This lets even non-coders approach software creation through a guided, conversational process, lowering the barrier from idea to working product.

Replit’s model-agnostic AI approach makes it easy for new users to adopt. It now integrates Anthropic’s Claude, OpenAI’s GPT-4o, and Google Gemini, routing tasks to the best performer. This is flexible, but Replit does not own the core AI models in its value chain.

Building a business, not just a product

Beyond the technical approach, Replit has, due to its long history, also established a clear path to profitability. The company moved away from targeting professional developers. Instead, it focused on average users, non-technical founders, product managers, designers, educators, and small business owners, who want to create software without writing code. That decision dramatically expanded its addressable market. To make the platform more accessible, it then adopted a usage-based pricing model tied to AI agent activity.

Enterprise clients like Zillow, Coinbase, Duolingo, HubSpot, and Mercedes-Benz pay per seat plus consumption fees. With gross margins increasing, Replit is growing quickly and sustainably. Partnerships with Microsoft and Google Cloud support this, enabling enterprises to purchase Replit through familiar channels such as the Azure Marketplace.

When the agent goes wrong

However, the rapid growth of Replit and its ability to enable non-technical users to develop software have also increased the risk for both the platform and end users.

In July 2025, an incident at the SaaStr conference highlighted the risks posed by autonomous AI development tools operating in live production environments. The episode drew attention to how these systems can act with broad system access while appearing to function normally, even as significant problems unfold in the background.

This raised a broad industry question: When autonomous AI agents control software, who is accountable if something goes wrong? Is it the user who granted access, the platform that built it, or the model provider?

Industry surveys show many developers have ethical or security concerns about AI-generated code. Many organizations say such code needs a thorough review.

Structural concerns add another layer of complexity. When businesses build entirely within Replit’s ecosystem, using its hosting, databases, and authentication tools, switching away becomes costly. This can create dependency on a single platform.

Rules for a world that moved too fast

To address these issues, policymakers are steadily moving towards a concrete framework for this new and evolving industry. Despite efforts in the U.S., the lack of federal laws specifically governing AI-assisted software development complicates the landscape.

While the U.S. lacks regulations, the AI Act in Europe imposes transparency obligations on foundation model providers, creating indirect compliance considerations for platforms like Replit. At the state level, emerging laws focus more on harmful AI use cases and disclosure requirements than on development tools themselves.

This ambiguity makes it difficult for platforms like Replit to navigate compliance across jurisdictions.

The bet behind the billions

Replit’s $9B valuation suggests investors are confident it can sustain the rapid growth it has demonstrated, increasing ARR from $2.8M to $265M over the past 18 months.

The optimism reflects a broader industry shift that could forever shatter the bottleneck in human-computer interactions. And if platforms like Replit are right about how software development will progress, then it is the infrastructure for that next expansion.

Writing became universal when tools faded into the background and the mechanics of ink, printing, and typesetting no longer posed barriers to expression. Replit is attempting something similar for software. If coding becomes as natural as writing a paragraph, the ability to build software may expand beyond the professional developer class. The real question is not whether AI can write code, but whether it can make software creation as effortless and as widely accessible as putting ideas into words.

Quick Bits, No Fluff

• Amazon AI backfires: Financial Times reports Amazon’s aggressive use of internal AI coding tools has triggered major outages and risky deployments, prompting stricter human review even as it cuts engineers.

• Perplexity vs Amazon: A Washington court temporarily blocked Perplexity’s AI shopping agent after Amazon alleged it scraped pages and overlaid affiliate links in violation of its rules and data policies.

• AI apps lose users: An AppsFlyer report finds AI-powered mobile apps monetize quickly but have about 30% higher annual churn than non-AI peers, showing novelty wears off fast.

Thursday Poll

Three Things Worth Trying

• Ship a tiny agent-built app: Pick a trivial internal tool, let an AI agent build it end-to-end, then do a hard line-by-line review of what you would change.

• Map your platform lock-in: List where your stack already depends on one vendor for hosting, auth, and databases, then write down what it would take to move.

• Set rules for non-dev builders: If non-engineers use tools like Replit, define clear policies for reviews, permissions, and what is never allowed to go live without an engineer.

The Toolkit

• Deepgram: Speech recognition engine that turns calls and recordings into accurate transcripts and searchable voice data.
  

• Descript: An AI video and audio editor that lets you edit by tweaking the transcript instead of a timeline.
  

• Drift: This AI chat platform turns website visitors into qualified conversations and booked meetings in real time.

Rate This Edition