When AI researcher Andrej Karpathy floated the phrase “vibe coding” in a 2025 tweet, describing a style of programming where you “give in to the vibes” and focus on intent rather than syntax, he crystallized a cultural shift already underway. His post went viral, resonating with developers willing to let AI assistants take the wheel.
Eight months later, that mindset has a poster child: Replit’s newly launched Agent 3. The company calls it “10× more autonomous” than prior versions, with the ability to test and fix code automatically and even orchestrate custom workflows. Unlike autocomplete-style copilots, Agent 3 aims to act more like a collaborator—capable of running extended development sessions with minimal oversight.
Beyond Autocomplete: Toward Autonomous Development
According to Replit, Agent 3 periodically tests applications inside a browser, identifies issues, and patches them on its own. CEO Amjad Masad has described internal runs where the agent worked independently for 4.5 hours straight, building an application that would normally take a week of traditional development. Replit emphasizes that its proprietary testing framework is more efficient than generic “computer use” models, though public benchmarks remain scarce.
The company frames this not as a smarter autocomplete, but as a shift in workflow: developers describe goals in natural language, and the agent does the heavy lifting. Failures still happen—Masad has admitted that prior generations often left non-technical users frustrated—but Agent 3’s architecture attempts to reduce those dead ends by creating multiple test environments and simulating different user behaviors.
The Testing Breakthrough—With Caveats
Traditional “computer use” models promised automated testing but often proved brittle and expensive. Replit’s answer is a targeted framework built for its environment. The company claims it’s 3× faster and up to 15× cheaper than the alternatives, though these figures come from Replit’s own benchmarks and should be taken as provisional. Ars readers will want to know how reproducible those numbers are outside lab conditions.
Notably, Agent 3 can replay browser sessions, giving developers visibility into what the AI actually did. That transparency could help address one of the biggest concerns with autonomous agents: trust.
Market Momentum: From Niche IDE to $3B Valuation
The numbers behind Replit’s trajectory are striking. Annual recurring revenue reportedly jumped from $2.8 million to $150 million in under a year, fueled by a community of over 40 million users. In September 2025, the company announced a $250 million funding round at a $3 billion valuation (as reported by Reuters and other outlets).
The broader AI coding tool market is also expanding rapidly. According to a recent market analysis, the AI code tool market was valued at USD 15.11 billion in 2025 and is projected to reach USD 99.10 billion by 2034 (AI Code Tool Market Size, Share, Trends and Analysis 2034). A separate industry guide noted that 97.5% of companies have already integrated AI into their software engineering processes (35 Best Vibe Coding Tools & AI Code Generators, 2025 Guide). While these numbers are impressive, forecasts of this kind should always be treated with caution.
Competitors: Cursor, Copilot, and the Platform Wars
Replit isn’t alone. Cursor has reportedly raised $900 million at a $10 billion valuation, with annual recurring revenue around $200 million. GitHub Copilot, backed by Microsoft, continues to dominate in enterprise integration. Amazon’s CodeWhisperer and Google’s Gemini-powered tools also loom large.
Masad argues that Replit’s decade of infrastructure work—its distributed file system with copy-and-write capabilities, ultra-fast environment forking, and multi-agent orchestration—provides a moat. He notes that “you can spin up a vibe-coding demo in an afternoon, but building a deep platform takes years.” That depth may help Replit differentiate, but whether it’s enough against hyperscale cloud providers remains an open question.
Should People Still Learn to Code?
Masad is blunt on this point: “If you want to build something, don’t wait to learn syntax. Start building. If you need to code, you’ll pick it up along the way.” It’s a provocative stance that clashes with decades of advice. He softens the argument by distinguishing between specialists (who still need to master code) and builders/entrepreneurs (who may be better off focusing on product and market).
That tension echoes a larger debate: Will AI coding tools democratize development or deskill the profession? Research suggests that tasks like UI component building and simple app scaffolding are already among the most disrupted. But high-end system design and debugging still demand human expertise.
Risks and Reality Checks
Autonomy comes with hazards. For every successful multi-hour autonomous run, there may be cases where oversight fails and the agent produces incorrect or unstable code. The challenge for Replit—and for all agentic coding tools—is ensuring robust verification and recovery mechanisms. Transparency, sandboxing, and user control will matter as much as raw capability.
This duality—massive productivity gains alongside unpredictable errors—defines the current state of agentic coding. Trust, verification, and safety nets will be critical as adoption scales.
Platform Risk and Model Agnosticism
Another uncertainty is dependence on foundation models. Replit currently leans on Anthropic’s Claude Sonnet 3.5 for agentic reasoning, OpenAI’s GPT-4 for code review, and Google’s Gemini for cost-sensitive operations. Masad stresses that loyalty is thin in developer tooling: “People will always switch to the best, cheapest, fastest option.” That pragmatism may shield Replit from model lock-in, but it also means competitors can pivot quickly.
The Future of Work
Masad doesn’t sugarcoat the implications: “People will lose their jobs.” Still, he frames it as economic churn rather than terminal decline. He points to the “communication tax” in large organizations—coordination overhead that slows progress—as a pain point AI could relieve. Early experiments show AI agents taking over internal workflows in areas like sales and HR.
But whether this creates new categories of work or hollows out mid-level engineering remains unresolved. For now, AI agents are best viewed as force multipliers with a long list of edge cases.
A Messy Leap Forward
Replit’s journey from a lightweight online IDE to a $3 billion AI platform highlights both the promise and perils of autonomous coding. Agent 3 shows genuine advances in autonomy and testing, but also exposes the fragility of AI-driven development. The “vibe coding” era may indeed be here, yet it’s messy, uneven, and still fraught with risk.
The best programming language may one day be English. But the best developers will likely remain those who not only talk to AI fluently, but also know when to double-check its work.
Unlock the Future of Business with AI
Dive into our immersive workshops and equip your team with the tools and knowledge to lead in the AI era.
