Replit launched Missions on November 17, a product that takes a goal from a team and assigns a swarm of AI agents through planning, building, testing, and revision until the work is finished or the money is gone. The pitch is delegation of projects that teams never staff: land enterprise customers, speed up a product, replace old software.

The flagship demonstration is DataCat, an application monitoring platform. According to Replit’s announcement, the company wanted to bring monitoring in-house to cut costs and gain control, so it handed Missions that goal. Two days and a $3,000 budget later, Replit says, agents had produced dashboards, alerts, and application tracing, plus DataCat’s own data-collecting agent and tracing libraries. Replit is also releasing a version others can host themselves, under an open-source license.

Every result below is Replit’s own, from controlled tests it designed and ran.

Replit says it planted a database bug and DataCat fired its alert after 38 seconds, inside a 60-second target, and pointed to the exact query and its logs. In a three-hour load test, the platform held 26 million application activity records plus 4.4 million log entries without losing any, and it cleared all 51 checks, 36 on data accuracy and 15 on alert reliability. In matched Node.js tests, Replit reports 39 percent more requests handled with 51 percent less monitoring-related processor overhead. Its agent used 82 to 108 MB of memory under load, versus 319 MB for the original platform’s agent, and made logs searchable in 0.5 seconds against 4.2.

Replit also pitted DataCat against a monitoring platform built from the same prompt with /goal in Claude Code. DataCat passed 51 of 56 API compatibility checks, versus 45 of 56. It took in 49,000 traces a second against 9,800, and 405,100 metric points a second against 224,600. The Claude Code build was faster at logs and covered more product areas, which Replit acknowledges. That is one run per setup, scored only on the product areas both builds shared, and Replit’s own notes say the tests do not show how DataCat compares with the original vendor’s hosted service or with Replit’s production systems.

The swarm claim is the one with the clearest mechanism behind it. On 13 NumPy benchmark tasks, Replit reports that under its Auto (Max) setting a single agent reached speedups of up to 4x, a 16-agent swarm reached 17x, and a 32-agent swarm reached 19x. Under Auto (Power), the figures were 2x, 7x, and 9x. The gains flatten between 16 and 32 agents, so doubling the swarm bought little extra. Replit also says Missions turned ¥100,000 into ¥4.82 million on E-Commerce Bench, a simulated year of running online stores, roughly 48 times the stake and more than three times the best score on the public leaderboard. That is a single run.

The controls matter as much as the benchmarks for anyone who would give agents real authority. A mission starts with an agreement that fixes a spending ceiling and names the actions that need a human’s approval. Replit’s example caps a sales mission at $1,000 and requires sign-off before agents contact prospects, share a proposal, edit customer-facing materials, or make product commitments. Users pick how often they hear back (immediately, daily, or weekly), and an independent AI judge scores the work against the success criteria agreed at setup. When money is exhausted before the criteria are met, the mission halts unresolved. Replit does not say how the judge is built or who audits it, and a grader that shares a vendor with the builder is a question buyers should ask.

Replit frames Missions as a step toward what it calls the Self-Driving Company, with people setting direction while agents handle execution. The company has not published pricing for Missions or said how often missions finish inside their budgets. Those two numbers decide whether a $3,000 monitoring build is typical or a showcase.

For a team weighing a build-versus-buy decision this quarter, the open-source DataCat is a free way to inspect what a two-day, $3,000 agent build produces, and the 5 failed API checks mark where to look first.

Replit, “Introducing Replit Missions,” Replit blog, November 17, 2026.