Introducing Replit Missions | Replit

Open original

Replit is launching Missions, a product where you define a goal and a budget and a swarm of AI agents plans, builds, and iterates toward it, with you approving consequential actions. The post's main proof is DataCat, an in-house application monitoring platform the agents built in two days for $3,000.

  • DataCat was built in 2 days within a $3,000 cap, and it caught a deliberately introduced database bug in 38 seconds against a 60-second goal.
  • Under a three-hour load test, DataCat stored 26 million activity records and 4.4 million log entries with no data loss, and passed all 51 data-accuracy and alert-reliability checks.
  • In Replit's own controlled comparison against a Claude Code build from the same prompt, DataCat ingested traces about 5x faster and metrics nearly 2x faster, though these tests were run by Replit and not on production systems.

Replit has launched Missions, a product where a team defines an outcome, sets a budget and approval rules, and then a coordinated swarm of AI agents plans, builds, tests, and iterates toward it. The examples are things like "win our first 10 paying enterprise customers" or replacing legacy software. Setup involves answering questions that pin down success, such as whether paid pilots count toward a customer total. Spending caps and approval gates are enforced: you might cap a mission at $1,000 and require sign-off before any prospect is contacted, any proposal goes out, or any product commitment is made. Progress updates arrive immediately, daily, or weekly, and an independent AI judge checks results against the criteria you agreed on. If the budget runs out first, the work simply stops with the outcome unresolved. The main proof point is DataCat, an application monitoring platform Replit built internally with Missions. The mission took two days and $3,000 to produce dashboards, alerts, application tracing, and a custom collection agent. In testing, DataCat caught a deliberately injected database bug in 38 seconds against a 60-second target and traced it to the exact query. Under a three-hour load test it stored 26 million activity records and 4.4 million log entries with no data loss, passing all 36 data-correctness and 15 alert-reliability checks. Its agent used 82 to 108 MB of memory against 319 MB for the original platform's agent, and logs became searchable in a median of 0.5 seconds rather than 4.2. A head-to-head against a build made from the same prompt with /goal in Claude Code showed DataCat accepting all four data types (logs, metrics, host data, traces) where the other accepted only logs, and ingesting traces about five times faster. Replit also reports a performance experiment on 13 NumPy benchmark tasks, where swarm size mattered a lot. A single agent in Auto (Max) mode reached up to 4× speedup, while 16 and 32 agents reached 17× and 19×. Auto (Power) followed the same pattern at 2×, 7×, and 9×. Read the numbers with some care, though. Replit notes that the DataCat tests were controlled and do not establish performance against the original platform's hosted service or against Replit's production systems. These are vendor-run experiments on a product Replit is selling, so the figures show what the tool can do under chosen conditions, not how it will perform on your own work.

Ask about this article