Replit Agent 3 tests its own apps in a real browser

Most AI coding tools spit out dozens of files and leave you to figure out why the submit button does nothing. Replit wants to take that headache off your plate entirely.

Replit just rolled out Agent 3, calling it their most autonomous system to date. Compared to Agent V2, the company claims it offers a 10x jump in autonomy. The headline feature is that it can spin up a real browser, click through the app it just built, catch broken flows, and fix the code on its own.

A self-testing developer in your browser

Instead of making you copy-paste console errors back into a chat prompt, Agent 3 uses a proprietary in-browser testing setup. Inside your workspace, you can literally watch the agent's cursor move across the preview pane, clicking buttons, submitting forms, and probing APIs.

It can even log into accounts using Replit Auth to verify sign-in flows. Replit notes that the agent does not run tests after every single prompt, but triggers them when it judges that enough code has changed to warrant a check. According to Replit, this custom test environment runs 3x faster and is 10x more cost-effective than using general Computer Use models.

There is also a new Max Autonomy mode in beta. When enabled, Agent 3 can run for 200 minutes or more without user input, maintaining its own to-do list and debugging along the way. You can track its progress live on your desktop dashboard or check in from your phone.

Replit also added a frontend-first creation path, letting you rapidly tweak UI layouts before deciding to spin up a full backend database.

Agents building agents

Agent 3 can also generate standalone agents and workflow automations. By selecting a dedicated option on the homepage, users can prompt their way into building Slack bots, Telegram bots, or scheduled cron jobs.

Instead of hunting down API secrets across developer portals, the agent launches guided connection flows for services like Notion, Linear, Dropbox, and SharePoint. It grabs credentials securely and gets back to writing the logic.

Use cases range from a Slack bot that searches your Notion databases to a pre-meeting assistant that runs web searches on calendar guests 20 minutes before a call and files summaries into Google Drive. The new model is available now for both free and paid users.

Why it matters

We think the AI coding race is moving away from raw code completion and directly into self-verification loops.

Building an app has never just been about generating syntax; it is about testing edge cases, spotting broken styling, and validating authentication. Replit bypassed expensive general-purpose computer-vision models to build a purpose-specific testing harness, which is a very smart move. It keeps generation costs reasonable enough to let everyday creators test functional prototypes without worrying about burning through tokens on basic UI clicks.

That said, giving an agent 200 minutes of continuous leash has obvious pitfalls. If the agent misinterprets your core data schema early on, it could spend three hours building deep logic on top of a flawed foundation. We recommend starting with tighter scope and checking its initial task breakdown before letting it run wild.

More broadly, Replit is signaling that it no longer sees itself purely as a code editor rivaling Cursor or Copilot. By blending UI creation, automated testing, and third-party SaaS integrations, it is taking direct aim at low-code workflow platforms like Zapier.