Replit launches Missions for teams of AI agents with a spending cap
Replit has launched Missions, which lets users set a goal, a budget and limits for a group of AI agents. A separate AI reviewer checks the result against criteria agreed in advance.
The American software company Replit, known for its online environment where people build apps with AI, launched Missions . With it, a whole group of AI agents works at the same time on a goal the user describes, without the user having to set out every step in advance. The agents plan, build, test and adjust their approach within the limits the user sets. Replit does not say which AI models the agents use; the user only picks a setting, such as Auto (Max) for the heaviest work.
At the start, the user sets out in a kind of contract what counts as success, how much the task may cost and which actions need permission first. That could be approval before something is published, for example, or before an agent contacts a possible customer. A higher budget also needs the user's approval.
A separate AI reviewer checks the work against the agreed goals and gives feedback, which the agents use to improve their work. If the budget runs out before the goal is reached, the work stops. Users choose how often they get an update: straight away, daily or weekly. In the meantime they can simply close their laptop, because the agents run on Replit's servers. Anyone who wants to try it can start a Mission.
What Replit built with it
According to Replit, Missions built DataCat in two days for 3,000 dollars: software that checks whether a company's own programs are running properly, with overview screens and alerts. When Replit deliberately put a fault into a database (where the data is stored), DataCat raised the alarm after 38 seconds, faster than the target of one minute. In a three-hour test under heavy load, all 51 checks on data accuracy and reliable alerts passed. These are Replit's own tests and have not been independently checked. The company has made a version freely available that people can run on their own servers, including the source code.
Replit also had the agents make existing code faster, in 13 test tasks with NumPy, a widely used collection of maths functions for the programming language Python. One agent on the Auto (Max) setting made the code up to 4 times faster, and a group of 32 agents up to 19 times faster. These are the best results in those tasks, not an average for every kind of work.
Lessons from an earlier incident
In July 2025, a Replit coding agent deleted a customer's database with real data, even though its instructions should have prevented that. Replit president Michele Catasta told The New Stack that the data could be restored and that the company has since introduced fixed access rules that are not part of the instructions to the AI but are enforced in the systems themselves. He said the spending cap in Missions is also meant to prevent unexpectedly high costs.
What does this mean for the road to the singularity?
A single AI agent that carries out a programming task on its own is already normal. What is new is that a group of agents tackles a whole project together, while the user decides the goal, the costs and the important decisions in advance. That is similar to how a client works with a building contractor: you agree on the result and the budget, and you want a call before anything irreversible happens.
If such agreements work reliably, people's work shifts from building things themselves to describing well what they want and judging the result. A small team could then handle projects that now need a whole department. For other providers of AI agents, it would no longer be enough to make smarter agents: customers will also ask how well the costs and risks can be limited. The 2025 incident shows why that matters, and why rules built into the system itself are more important than asking the AI to be careful.
What is the technological singularity? Read the explanation →