Replit Introduces Agent 3, Its most autonomous AI agent to date

Replit has announced the launch of Agent 3, the third iteration of its artificial intelligence agent for software development. This version presents substantial improvements over Agent V2, with an autonomy capacity ten times greater according to company data.

The main innovation of the system is its ability to automatically test the applications it builds. Agent 3 uses a real browser to verify the functionality of buttons, forms, APIs, and data sources, automatically correcting detected issues. This proprietary testing system from Replit operates three times faster and is ten times more cost-effective than existing Computer Use models on the market.

The tool offers two operating modes. The application testing mode is enabled by default and allows the agent to periodically verify app functionality during development. The Max Autonomy option, available in beta, enables Agent 3 to operate for extended periods with minimal supervision, managing complex tasks independently.

Another notable addition is Agent 3's capability to create other agents and automations. Users can develop bots for Telegram or Slack, time-based automations, or integrations with applications like Notion, Linear, Dropbox, or Sharepoint using natural language. The system guides users through simplified connection flows, eliminating the need to manually search for and copy API keys.

Agent 3 can run autonomously for up to 200 minutes, building, testing, and fixing applications with minimal manual intervention. Users can monitor their project progress in real-time from the web or mobile devices.

Replit has also redesigned the application creation flow. Users can now choose between developing a complete application from the start or beginning with a frontend version for rapid prototyping, adding the backend later if needed.

The service is available for both free and paid users of the Replit platform.

Key points

• Agent 3 multiplies autonomy capacity tenfold compared to Replit's previous version
• Incorporates automatic testing in real browser that verifies and corrects errors autonomously
• The system operates three times faster and ten times more cost-effectively than Computer Use models
• Enables creation of other agents and automations using natural language
• Can run autonomously for up to 200 minutes without manual intervention
• Includes simplified integrations with applications like Notion, Linear, Dropbox, and Sharepoint
• Offers two modes: default application testing and Max Autonomy for extended complex tasks
• New creation flow allows choosing between full-stack development or initial frontend