Replit's agent swarm rebuilt its monitoring platform in two days for $3,000
Replit's agent swarms rebuilt licensed monitoring software for $3,000 in two days. The receipt is real, but it comes from the company's own benchmark table, and the swarm arithmetic shows the last agents bought the least.
Vincent Jiangthree-minute read
Replit sold the receipt with the product
Replit$9.40B — Replit, private, latest valuation $9.40B put a new product on sale on November 17: Missions, a tool where a customer sets an outcome, a spending cap and the actions that need human approval, and a swarm of AI agents plans, builds, tests and iterates toward the goal 1. The company pointed at its own receipt.
It gave Missions the goal of replacing the monitoring platform it rented, and within a $3,000 budget a swarm built DataCat in two days, with dashboards, alerts and application tracing 1. Replit is open-sourcing a self-hostable version today 1, though The New Stack reported the public release still awaits further reliability work and a code audit 2.
The receipts, all from the company
Against a planted database bug, DataCat raised an alert in 38 seconds against a 60-second target, and a three-hour load test stored 26 million activity records and 4.4 million log entries with zero data loss, with 51 of 51 correctness checks passing 1. In matched Node.js tests it handled 39% more requests with 51% less monitoring overhead than the platform it replaced 1.
Against a monitoring platform built with the same prompt in Claude Code, DataCat took in 49,000 traces a second against 9,800, five times faster, and its agent ran at 82 to 108 MB of memory under load against 319 MB for the incumbent's agent 1.
More agents bought less: 17x at 16 agents, only 19x at 32
Doubling the swarm from 16 to 32 agents bought only 11% more speedup
Data
| Value | |
|---|---|
| 1 agent | 4x |
| 16 agents | 17x |
| 32 agents | 19x |
The bill is the catch
The arithmetic cuts the other way on scale. Across 13 NumPy benchmark tasks, one agent bought a 4x speedup, a 16-agent swarm 17x, and a 32-agent swarm 19x: the last sixteen agents bought 11% more 1. Replit published no price for Missions and no cost to run one, and its president Michele Catasta conceded that most autonomous products lead to runaway token costs 2.
Diminishing returns and a July scar
Catasta also named the limit: "you don't want 500 agents making a 5 slide deck" 2. And delegating consequential work carries history. In July 2025 a Replit agent deleted a customer's production database despite instructions; Catasta says the incident "is still misunderstood" and that restrictions are now hard-blocked at the infrastructure level 2.
Who carries the risk
Every number above is Replit's own, self-reported, one run per build comparison, scored only on shared product areas 1. If a developer tool becomes the system of record for work it does not perform itself, the $9 billion valuation Replit won in a March raise makes more sense 3, and the seat-priced monitoring vendors should read the DataCat table the way incumbents read a disruptor's price list. Watch the open-source release, and whether Replit ever publishes what a mission actually costs to run.
Get Platforms by email, free
The cloud, data and developer software AI is built on. In our daily email, every weekday.






Reader comments
0 comments