
Image: Replit
Why it mattersA buyer sees a cap and a success test before the agents start. A team sees a built-in rule for what the swarm cannot do without a human.
A single coding agent you have to watch is a hard thing to sell to a team. Replit's new product is the opposite shape. You write down what you want and what the agents are not allowed to do. The swarm runs until it meets the goal or the money runs out.
The Replit Team announced Missions on November 17. It coordinates many agents against one goal and gives the buyer three dials before they start. A budget cap that stops the work when it runs out. A list of actions an agent must ask about first. And a success test an independent AI judge checks at the end.
What a Mission is
A single agent is one worker. A Mission is a shared job that a swarm of agents plans, builds, tests and iterates on.
The approval list covers the actions a buyer is most nervous about letting an AI do. Contacting a prospect. Sending a proposal. Changing a customer-facing page. Making a commitment on a product. Each can be set to "always ask first".
The notification settings are plain: get told about blockers, progress and status, as they happen or in a daily or weekly digest. The success criteria are written in plain words during setup. An independent judge reads the output against them before Replit marks the mission done.
Replit has not published a price. The post points enterprise buyers at sales.
The one example Replit shows
Replit built an app monitoring product called DataCat as a Mission run by its own team. The team set a cap of $3,000 and gave the swarm two days.
The build has a monitoring agent, tracing libraries, dashboards and alerts. It also has two storage engines, one on Postgres and one on ClickHouse. DataCat's own tests caught a seeded database bug in 38 seconds, against a 60 second target. It processed 26 million activity records and 4.4 million log entries in three hours. No data was lost. All 51 accuracy and reliability checks passed.
Replit also published comparisons with an existing commercial platform and with a Claude Code build on the same job. DataCat handled 39 percent more requests with 51 percent less processor overhead than the existing platform. It took in 49,000 traces per second, five times the Claude Code build. And 405,100 metric points per second, nearly twice. The monitoring agent used 82 to 108 MB of memory against 319 MB. Log search answered in 0.5 seconds against 4.2.
The post is clear that these are controlled tests, not production numbers. They used matched workloads, one run per setup, and older reference versions. The Claude Code comparison only read logs, where DataCat also handles metrics and traces.
The DataCat code is open-source.
For a team buying this
The useful part of this release is not the swarm. It is the three dials around it. Set a cap and the money is the ceiling. Set an approval list and the AI cannot send the customer email it wants to send. Set the success test and somebody reads the result against it before calling the job done. Those are the controls a buyer needs to put an unattended AI anywhere near work that matters.
Source
The Replit Team, "Introducing Replit Missions," Replit. DataCat is Replit's own test build, not a customer case.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

