Replit Missions Hands a Goal and a Spending Cap to an Agent Swarm, and an AI Judge Decides When the Job Is Done
Replit launched Missions on November 17, 2026, letting users set an outcome, success criteria, a budget and approval boundaries for a swarm of cloud agents. Replit says Missions built DataCat, an application monitoring platform it has now open-sourced under Apache-2.0, in two days on a $3,000 budget. Pricing by plan was not disclosed.
Replit’s newest agent product starts with an agreement rather than a prompt. On November 17, 2026, the company launched Replit Missions, a product in which a user defines an outcome, agrees on how success will be measured, sets a spending limit and lists which actions need human sign-off. Replit then coordinates a swarm of agents to plan, build, test and iterate toward that outcome until the goal is met or the money runs out.
The company pitches Missions at work teams “haven’t had time or capacity for,” with examples such as winning more enterprise customers, improving product performance or replacing legacy software. It describes the product as a step toward what it calls the “Self-Driving Company,” in which people set direction and stay accountable while agents carry out execution and coordination.
A contract before any code
A mission starts from a conversation. Users can type /mission, accept a suggestion from Replit’s agent, or pick “Mission” from the New menu in the sidebar. Rather than asking for a detailed plan, setup asks questions that pin down success. For a goal such as “win our first 10 paying enterprise customers,” Replit’s example asks what counts as an enterprise customer and whether paid pilots count toward the 10.
The answers go into a mission agreement that sets the spending limit, the actions that require permission and the tradeoffs the agents may make, such as offering discounts on annual contracts. In Replit’s example, a mission capped at $1,000 must get approval before contacting prospects, sharing a proposal, changing customer-facing sales material or making product commitments. Replit says those limits stay in place as the approach evolves.
Users can also choose which connectors a mission can access. On engineering work, Replit says changes arrive as pull requests and reviewable files, and anything outside the agreed boundaries waits for a person.
Swarms, a judge and a hard stop
Once running, agents work in parallel in the cloud, so users can close their laptops. In the sales example, one agent researches buyers while another builds a demo; after approved outreach is sent, the mission reviews which accounts respond and shifts effort accordingly. Users choose whether to hear about blockers, progress and status immediately, in a daily digest or weekly.
Quality control sits with a separate agent. Replit says an independent AI judge checks results against the agreed success criteria, and the working agents use its feedback to refine their output. A mission is marked complete when the criteria are met. If the budget runs out first, work stops with the outcome unresolved. Messages a user sends during an active mission count toward the budget, and users can leave mission mode at any time.
Replit offered one internal measure of why it uses swarms. Across 13 NumPy performance tasks, a single agent on its Auto (Max) setting achieved speedups of up to 4x, a 16-agent swarm reached 17x and a 32-agent swarm reached 19x. Auto (Power) showed the same pattern at 2x, 7x and 9x.
DataCat, built in two days and now open source
The headline demonstration is DataCat, an application monitoring platform Replit built to bring monitoring in-house. Replit says a mission produced dashboards, alerts, application tracing, a collection agent and tracing libraries in two days within a $3,000 budget.
Replit’s tests found that DataCat raised an alert 38 seconds after engineers injected a database bug, against a 60-second target, and handled 26 million activity records and 4.4 million log entries over a three-hour load test with no data lost. All 51 data-correctness and alert-reliability checks passed.
Replit also compared DataCat with a monitoring platform built from the same initial prompt using /goal in Anthropic’s Claude Code. DataCat passed 51 of 56 API compatibility checks against 45 for the Claude Code build, ingested 49,000 traces per second against 9,800, and handled about 405,100 metric points per second against 224,600. The Claude Code build covered more product areas and took in logs faster. Replit notes that the comparison used one run per setup and that its tests do not establish performance against the original platform’s hosted service.
Replit has released a self-hostable version as replit/datacat on GitHub under the Apache-2.0 license. The repository describes it as self-hosted observability for metrics, logs, traces, dashboards and alerts, storing data in PostgreSQL or ClickHouse and accepting OpenTelemetry data.
On E-Commerce Bench, where an agent runs simulated online stores for a year starting with ¥100,000, Replit says one Missions run ended with ¥4.82 million, about 48 times the starting money and more than three times the best score on the public leaderboard.
Replit has not said which plans include Missions or how mission work is priced beyond the user-set cap. Its FAQ says only that work performed during a mission counts toward the budget set upfront.
- #Replit
- #Replit Missions
- #Agent Swarms
- #Multi-Agent
- #Open Source
Author
Raj M
Contributor
AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.