On November 17, 2026, Replit launched Missions: users only set a goal, a budget, and boundaries, while a swarm of agents working in parallel handles the planning, building, testing, and iterating. A separate AI judge then checks the result against criteria agreed in advance — the mission counts as complete only when the criteria are met, and if the budget runs out first, work stops. This is not about opening a few more agent windows; it turns delegation itself into a product.
The process runs in four steps. First, set the goal — for example, 'win our first 10 paying enterprise customers' — without having to decide upfront which industries to target or which buyers to approach. During setup, Missions asks questions such as 'what counts as an enterprise customer?' and 'do paid pilots count?', pinning down the definition of success first. Next, set the budget and boundaries: agree how much to spend and which actions need your approval first, such as contacting prospects or changing customer-facing sales materials. Then the agent swarm works in parallel — one researches customers while another builds a demo — adjusting direction based on feedback, reporting progress at the frequency you choose, and coming back to you when it hits a blocker. Finally, an independent AI judge checks the result against the agreed criteria, marks the mission complete when they are met, and keeps iterating when they are not.
Replit first used Missions on itself to build DataCat, an application monitoring platform: in two days and on a $3,000 budget, it produced dashboards, alerts, and request tracing. When the team deliberately planted a database bug, DataCat raised an alert in 38 seconds, beating its 60-second target. A three-hour load test stored 26 million activity records and 4.4 million log entries with zero loss, and all 51 data and alert checks passed. The same mechanism grew ¥100,000 of starting capital to ¥4.82 million in the E-Commerce Bench business simulation — about 48 times the starting amount, more than three times the best score on the public leaderboard. A self-hostable version of DataCat is now open source.
One caveat: these numbers all come from Replit's own controlled tests, not independent benchmarks. Agent swarms genuinely get faster with more agents on parallel work like NumPy optimization — 32 agents reached up to 19× speedup — but on work where the goal is vague and success is hard to measure, whether the judge works well and whether the budget gets burned for nothing is something you have to try for yourself. Missions is now available from the 'New' menu in the Replit sidebar. Replit Agent could already run errands for you — see Replit Agent's new abilities: buying supplies for your company and placing Google ads on its own.