Replit’s Agent 4 coordinates multiple AI agents to build apps in parallel

AI coding tools can generate code quickly, but managing large changes across a project remains a challenge. Replit’s latest agent release attempts to organise that work into discrete tasks that can be tracked, reviewed, and executed in parallel.

Founded in 2016, Replit was originally known for its browser-based coding environment, where developers could write and run code entirely in the cloud. More recently, the San Francisco-based startup has shifted its focus toward AI-assisted software development.

That shift continues with Agent 4, the newest version of Replit’s AI system designed to build applications from natural-language instructions.

According to the company, Agent 4 can break a request into smaller jobs and run them concurrently while maintaining a shared understanding of the project.

Those tasks appear as individual items in a project board, where developers can review proposed work — such as connecting a database, adding authentication, or designing a landing page — before approving the agent to proceed. Once approved, the agent moves those tasks through stages of execution until they are ready to be merged into the application.

The idea is to help developers manage complex application builds that involve many interacting components rather than isolated code snippets.

The launch follows a series of updates aimed at organising how Replit’s agents operate inside a project.

One example is Plan Mode, which keeps the AI agent in a planning role until a developer approves the proposed changes. Instead of modifying files immediately, the system first outlines the work it intends to perform.

Another feature, known as Queue, allows developers to stack multiple instructions while the agent processes earlier tasks. The tool then executes those requests sequentially rather than forcing users to wait for each step to finish.

Both features address a similar challenge: how to bring structure and oversight to automated coding systems that can make large changes to a codebase.

Agent 4 extends that direction by coordinating multiple tasks across a project at the same time.

From single-prompt tools to coordinated agents

Earlier versions of Replit’s AI agent focused largely on autonomy. Agent 3, introduced last year, could generate code, run tests, and attempt fixes on its own after receiving a prompt.

Agent 4 places more emphasis on how multiple pieces of work are organised inside a project.

Instead of running a single task from start to finish, the system can coordinate several tasks within the same build. Those tasks move through stages of planning, approval, and execution before being merged into the project. This might involve generating a front-end interface while setting up database schemas and authentication logic in parallel.

A key part of this approach is the use of parallel agents that can work on different tasks simultaneously while maintaining a shared view of the project.

The release also introduces a visual design canvas where developers can experiment with interface layouts before applying those changes to the underlying codebase.

‘Tools for professional developers, and tools for everyone else’

Peter Yang, Roblox product lead and author of the Behind the Craft newsletter, spent time testing Agent 4 ahead of its public release. Yang said he spent the weekend building a habit-tracking app, experimenting with the new infinite canvas — a workspace for exploring and comparing multiple interface variants inside the project — alongside its parallel agents. In his walkthrough, Yang shows how different UI ideas can be explored on the canvas while agents work on other tasks at the same time, with work moving through a board as it progresses from draft plans to active builds and then ready for review.

While Yang was broadly positive about the experience, he also used the experiment to reflect on where tools like Agent 4 may sit within the rapidly evolving AI coding landscape.

Indeed, Yang argues that the AI coding market is now split into two broad segments – “tools for professional developers and tools for everyone else.”

“For developers, I think it’s a two-horse race between Claude Code and OpenAI’s Codex,” Yang says. “These tools are powerful, fast, and deeply integrated into model and developer workflows.”

Platforms such as Replit, however, are positioning themselves as tools for the latter category.

According to Yang, Replit is making two key bets: that non-technical users will increasingly collaborate directly with AI agents, with designers generating ideas on a canvas and product managers coordinating work through task boards; and that code itself is becoming the foundation for many forms of knowledge work, enabling agents to generate not only applications but also presentations, animations, and other digital outputs.

“I think Agent 4 is a real step forward on both fronts,” Yang writes. “The canvas could use a bit more polish, but the direction is clear. We’re going to work with AI agents to build all kinds of things in the future, not just apps.”

A growing market for AI development tools

Yang’s assessment also reflects the broader momentum behind tools designed to automate more of the software creation process.

Indeed, Replit has just raised $400 million in new funding, valuing the startup at roughly $9 billion and signalling growing investor interest in tools designed to automate software creation.

Agent 4 doesn’t remove the need for engineers to review architecture or validate code changes – a point underscored by recent outages at Amazon linked to AI-assisted coding tools. But it illustrates how AI coding systems are moving beyond simple assistants toward tools designed to coordinate substantial parts of an application build.

For developers experimenting with these tools, the question now is how much responsibility they are willing to hand to automated agents inside their projects.