Replit says its agent swarm built Radar in 2 days for $3,000

Replit has launched Missions, a product that hands a large goal to a group of AI agents working in parallel. You set the goal, a budget and the actions that need your approval, and the agents can then work on it for hours, days or even weeks. Before the launch, Replit used Missions to build Radar, its own app monitoring platform. The company says the agents built it in two days for $3,000, with Replit engineers giving direction. The details come from Replit’s president, Michele Catasta, in an interview with The New Stack.
Groups of agents like this, often called agent swarms, have become a race among AI companies. In the past week alone Google and Cursor showed competing approaches to coding with several agents at once. Google’s version is a /swarm command in its Antigravity App that coordinates subagents over a shared message board. Anthropic already has an experimental Agent Teams feature in Claude Code.
How a mission works
You type /mission in a conversation and describe the goal in plain English, for example getting every page of a website to load in under 500 milliseconds. Replit then sets up what it calls a Mission Agreement, which records the success criteria, the spending limit and the actions that need a human to approve them.
You don’t have to write a plan. Catasta says that’s on purpose, because a step-by-step plan “would force the human to worry about implementation details, and not afford the flexibility the AI system might need to accomplish complex objectives.” The agents work out how to get there and change course as they go. A separate agent then checks the result against the success criteria and sends it back for more work if it falls short. Only then does it reach you.
Each agent works in its own isolated cloud environment, with access to tools, databases, logins and deployment. Agents edit separate copies of the project, and when their changes overlap, another agent settles the conflict. Replit first built this setup so that human developers could work on the same project together. Catasta also credits the AI models themselves: swarms, he says, “have only started to be useful over the last few months” as models got better at splitting up work.
Keeping costs and agents in check
A mission can’t start without a spending limit. Replit checks spending against it at every step and blocks paid services once the limit is reached. “Most autonomous or multi-agent products lead to runaway token costs,” Catasta says.
More agents don’t always help. “You don’t want 500 agents making a 5 slide deck,” Catasta says, and Replit suggests how many agents each job should get.
The bigger worry is whether agents stay within the limits you give them. In July 2025 a Replit coding agent deleted a customer’s live database despite instructions limiting what it could do. Catasta says the incident “is still misunderstood,” because the database could be restored with Replit’s recovery tools. Since then, he says, Replit has added clearer rules on which integrations agents can use and which actions need permission, and it enforces them at the infrastructure level. “Anything not allowed is hard blocked by our proxy service.”
What Replit built with it
According to Replit, the Radar agents produced dashboards, alerts and tracing, which follows a request as it moves through an app. In a test, Replit deliberately broke a database. Radar spotted the fault and raised an alert in 38 seconds, inside its 60-second target. Replit also says that using 32 agents at once made some workloads in NumPy, the widely used Python library for numerical computing, run up to 19 times faster.
Radar was built for Replit’s own use, but the company plans to release it as open source after more reliability work and a code audit.
Up next
CNBC: Anthropic IPO likely to price above $70-78
Anthropic’s IPO looks set to price above its $70-78 range, multiple sources told CNBC, with order books closing Tuesday and Nasdaq trading set to start Nov. 19. What price per share do you think Anthropic’s IPO lands at?
Related stories
- Antigravity’s /swarm runs subagents over a message board
- Cursor’s new coordinator agent writes no code itself
- Model routing becomes the default across Replit accounts
- Claude Code stops wiping .env files
- Perplexity puts Claude, GPT and Gemini on one team
- Microsoft rolls out Autopilot agent in Frontier preview
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

