Replit has launched Missions, a product that hands a goal to a swarm of AI agents and lets them plan, build, test and iterate toward it. The user sets the outcome, the spending limit and the boundaries, and the swarm works inside them.
How a mission works
Each mission starts with a setup conversation that defines success: what counts as an enterprise customer, whether paid pilots count, or whether success means signed annual contracts above a minimum value. Those answers shape which accounts the mission prioritises and how results are measured.
The mission agreement then sets the spending limit and which actions need approval, plus the trade-offs agreed in advance, such as offering paid pilots or discounts on annual contracts. A user might cap spending at $1,000 and require approval before the agent contacts prospects.
An independent AI judge checks results against the success criteria agreed at the start, and the agents use its feedback to refine their work. A mission is marked complete when those criteria are met, or stops unresolved if the budget runs out first.
The DataCat self-test
Replit used itself as the first example, asking Missions to build DataCat, an application monitoring platform it could run on its own infrastructure. In two days, within a $3,000 budget, a swarm built dashboards, alerts and application tracing, then tested the platform against measurable goals.
To test whether DataCat could diagnose a failure, the team deliberately introduced a database bug. DataCat raised an alert in 38 seconds, beating its 60-second target, and traced the failure to the exact database query and its logs in the test run.
The mission also ran a three-hour load test in which DataCat stored 26 million records of application activity and 4.4 million log entries with no data lost. All 36 data-correctness checks and 15 alert-reliability checks passed, and the platform is now open-sourced.
Our Read
Every performance figure here is Replit’s own, from controlled tests the company ran itself. The blog says so plainly, including that the Claude Code comparison used one run per setup. Treat the numbers as a launch claim, not as independent proof of what swarms can do.
What is new is not the benchmark but the control model. Budget caps, approval checkpoints and an independent judge address the two things that stop teams trusting autonomous agents in production: runaway cost and unsupervised action. That is the launch’s real claim.
Why It Matters
If Missions works as described, a small team could take on work it keeps postponing: the product launch, the refactor, the software it wants to build in-house. The budget cap turns an open-ended agent experiment into something a manager can actually approve.
Background
Replit describes Missions as a practical step toward what it calls the self-driving company, where people set direction and remain accountable while agents carry out the execution. Missions is available now from the sidebar, or through Replit’s sales team for organisations.
The swarm angle is not new to Replit watchers: its 32-agent swarm hit 18.9 times the speed of a single agent on NumPy tasks, a benchmark covered in detail this week. Missions is the product that puts that swarm to work under a budget.
Source: Replit blog


0 Comments