Replit on November 17 released Missions. Instead of instructing an agent step by step, users describe the outcome they want and set a budget and permissions. Replit then coordinates a group of agents to plan, build and test, closing in on the goal round by round. Alongside it, Replit open-sourced DataCat, an application monitoring platform that was itself built with Missions.
It is Replit's second agent move this month. Last week it let Replit Agent make purchases and run Google ads on a user's behalf; this time it swaps a single agent for a whole team.
How a mission runs
The example in Replit's blog post is “land the first 10 paying enterprise customers.” During setup, Missions first asks what counts as success, for instance whether a paid pilot qualifies.
Next come the budget and approvals. In the example the cap is $1,000, and a human must sign off before the system contacts prospects, sends proposals, edits sales materials or makes commitments about the product. Once approved, several agents work in parallel: one researches buyers while another builds a demo. After outreach goes out, the system looks at which messages led to real conversations and adjusts the next round.
Progress reports can be set to instant, a daily digest or weekly, and the agents run in the cloud, so closing the laptop does not stop them. When the mission ends, a separate AI reviewer checks the result against the criteria agreed in advance. The blog also describes the other possible ending: if the budget runs out first, the work stops and the outcome is left unresolved.
DataCat and the comparison build
DataCat is a monitoring platform Replit built for itself with Missions, with dashboards, alerts and tracing. It took 2 days and a $3,000 budget. The open-source release is a self-hostable version, with the code at replit/datacat on GitHub; the blog does not mention a license.
Replit shared a set of numbers from controlled tests:
- With a database bug injected on purpose, DataCat raised an alert in 38 seconds, against a 60-second target;
- In a 3-hour stress test it wrote 26 million activity records and 4.4 million log entries with zero loss, and passed all 51 data-correctness and alerting checks;
- Against a comparison build made with Claude Code, trace throughput was 49,000 versus 9,800 per second, metric writes 405,000 versus 225,000 per second, and median log search time 0.5 seconds versus 4.2 seconds.
Several things about the comparison are unclear. The blog does not say how much budget or time the Claude Code build used, or what prompts it was given. Replit itself also states that these tests do not show DataCat is better than the commercial managed platforms it would replace, nor do they represent performance in Replit's production environment.
Do more agents mean faster?
Replit also published a scaling test on 13 NumPy tasks, looking at how agent count relates to speedup. On the Max tier, 1 agent gives up to a 4x speedup, 16 agents 17x and 32 agents 19x; on the Power tier the three figures are 2x, 7x and 9x.
Going from 16 to 32 agents doubles the headcount but adds only about two times more speedup, so the gain from parallelism has already flattened noticeably in that range.
Compared with products such as Cursor and Devin, which also push background agents, what sets Missions apart is that budget caps, approval checkpoints and AI acceptance review are built into the task definition, and the target extends from writing code to business outcomes such as customer acquisition. The blog does not say how Missions is priced or how many parallel agents each plan allows; enterprise users are simply told to contact sales.
Sources: Replit official blog, CocoLoop, DataCat GitHub repository; performance figures come from Replit's controlled tests, and the build conditions of the comparison version have not been disclosed.