Replit launches Agent 3 with 3-hour autonomous coding

If you have tried building anything with AI coding tools lately, you know the routine: it generates a slick-looking prototype in two minutes, hits an obscure dependency error, and leaves you to clean up the mess for the next two hours.

Replit thinks it just cracked that bottleneck. CEO Amjad Masad announced the rollout of Agent 3, calling it the full self-driving moment of software development.

Alongside the release, Replit confirmed a fresh $250 million funding round that triples its valuation to $3 billion. The company's financials tell an even wilder story: annual revenue skyrocketed from $2.8 million to $150 million in less than a year, powered by 40 million users and enterprise accounts like Zillow and Duolingo.

Surviving three hours without a babysitter

Most coding assistants today are essentially fancy autocomplete engines. They can whip up boilerplate or script a basic landing page, but shipping production software requires testing, debugging, and refactoring—the exact places where agents typically hallucinate or stall out.

Replit measures autonomy by uninterrupted runtime. While Agent 2 could run for roughly 20 minutes before needing a human to unstick it, Agent 3 can operate for over 200 minutes straight without any human intervention.

During those three-plus hours, the agent writes code, runs the environment, spots broken components, and fixes bugs autonomously, effectively mimicking a software engineer's iterative loop.

Masad noted that this jump did not come from waiting on a smarter single model. Instead, Replit shifted to a multi-agent architecture where distinct AIs divide and conquer: one handles high-level project planning, another writes the implementation, and a separate model tests and verifies the output.

Paul Roetzer, CEO of the Marketing AI Institute, pointed out on a recent podcast that this approach effectively outpaces standard scaling benchmarks like METR's seven-month rule by engineering around individual model limits.

Why it matters

We think the real leap here is psychological as much as technical. Up until now, using AI agents felt like managing an intern who required a check-in every quarter-hour. If an agent can genuinely work alone for three hours, you can hand off entire feature tickets before stepping into a meeting or stepping away from your desk.

For the developer tooling market, this shifts the battle away from raw model benchmarks and toward agentic endurance. Competitors like GitHub Copilot and Cursor will be under immediate pressure to prove their agents can handle prolonged, multi-step execution rather than just quick in-editor prompts.

Still, there is a clear trade-off. Giving an AI free rein for three hours means accepting massive context drift if things go sideways early on. If the planning agent makes a bad architectural bet in the first twenty minutes, you might return to hundreds of lines of hallucinated fixes built on top of a flawed foundation. It is definitely worth testing on standalone prototypes right now, but we would hesitate before letting it refactor a core production repo unchecked.