Replit launches Missions: agent swarms work toward goals within budgets and approval limits you set
Replit has launched Missions, a product where users define an outcome and Replit coordinates a swarm of agents to plan, build, test and iterate within a budget and boundaries the user sets. Replit says it used Missions to build an application monitoring platform, DataCat, in 2 days within a $3,000 budget.
Replit has launched Missions, a product in which a team defines an outcome and Replit coordinates a swarm of agents to plan, build, test and iterate toward it within a budget and boundaries the team sets. Replit positions it for complex work teams lack capacity for, such as winning enterprise customers, improving product performance or replacing legacy software.
How Missions works
During setup, Missions asks questions to define success. In Replit's example mission, "win our first 10 paying enterprise customers," those questions include what counts as an enterprise customer and whether paid pilots count toward the first 10 or only signed annual contracts above a minimum value.
The Missions agreement sets the spending limit and which actions require permission. Replit's example caps spending at $1,000 and requires approval before the Agent contacts prospects, shares a proposal, changes customer-facing sales materials or makes product commitments. Replit says these limits and approval requirements remain in place as the approach evolves.
Agents work in parallel, for example one researching buyers while another builds a demo. Users choose which updates they receive (blockers, progress, general status) and how often: immediately, in a daily digest or weekly. An independent AI judge checks results against the agreed success criteria, and agents use its feedback to refine their work. A mission is marked complete when the criteria are met. If the budget runs out first, work stops with the outcome unresolved.
Swarm scaling results
Replit tested swarm size on 13 representative NumPy benchmark tasks, where agents explored alternative implementations, optimized memory access patterns and validated correctness automatically.
| Configuration | 1 agent | 16 agents | 32 agents |
|---|---|---|---|
| Auto (Max) | up to 4× | 17× | 19× |
| Auto (Power) | 2× | 7× | 9× |
These are Replit's own results.
DataCat: Replit's internal test case
Replit says it gave Missions the goal of building DataCat, an application monitoring platform it can run and customize on its own infrastructure. According to Replit, Missions built dashboards, alerts, application tracing, DataCat's own agent for collecting server data, and tracing libraries in 2 days within a $3,000 budget.
Replit reports the following results:
- Failure detection: with a deliberately introduced database bug, DataCat raised an alert in 38 seconds against a 60-second goal and traced the failure to the exact query and its logs.
- Load test: over three hours, DataCat stored 26 million application activity records and 4.4 million log entries with no data lost. All 36 data-correctness checks and 15 alert-reliability checks passed.
- Node.js efficiency: in matched tests, the application handled 39% more requests with 51% less monitoring-related processor overhead.
- Agent memory: 82–108 MB for DataCat's agent versus 319 MB for the original platform's agent (240-second average).
- Log searchability: 0.5 seconds with DataCat's agent versus 4.2 seconds with the original platform's agent (median).
- Storage engines: DataCat supports two, Postgres and ClickHouse.
Comparison with a Claude Code build
Replit also compared DataCat with a monitoring platform built from the same initial prompt using /goal in Claude Code. DataCat accepted all data types from the original platform's agent (logs, metrics, host data and traces); the Claude Code build accepted logs only. DataCat passed 51 of 56 API compatibility checks on each storage engine versus 45 of 56 for the Claude Code build. Trace intake was 49,000 traces/second with DataCat (Postgres) versus 9,800, and metric intake was 405,100 points/second versus 224,600. The Claude Code build, which used SQLite as its single storage engine, covered more product areas and ingested logs faster.
Replit's caveats: the controlled tests do not establish performance against the original platform's hosted service or on Replit's production systems, and the build comparison used one run per setup.
What this means
Missions packages agent orchestration around three controls: an agreed definition of success, a hard budget, and approval gates on outward-facing actions. Those controls address the main operational risks of long-running autonomous work, namely unbounded spend and unsanctioned actions. The completion rule also matters: an independent judge evaluates against criteria the user agreed to, and a mission that exhausts its budget stops unresolved rather than being declared done.
The NumPy results are Replit's own, but they show a consistent pattern across both configurations: more agents in the swarm produced larger speedups. The DataCat figures come from a single internal project with one run per setup in the Claude Code comparison, so they are best read as a case study rather than a general benchmark.
Related Articles
Microsoft makes Word, Excel and PowerPoint in Copilot generally available, calls Copilot "the new OS for work"
Microsoft CEO Satya Nadella announced Tuesday at Ignite the general availability of Word, Excel and PowerPoint inside Copilot, which he called "the new OS for work." The company also introduced Copilot Autopilot, Microsoft IQ and expanded Agent 365 controls, with pricing for Autopilot still unannounced.
Perplexity Brain Can Now Create and Update Agent Skills; 80,000+ Personal Skills Written for 20,000+ Users
Perplexity said on November 18, 2026, that Brain, the personalization system for its Computer agent, can now create and update reusable agent skills. According to the company, Brain wrote more than 80,000 personal skills for over 20,000 users between August 1, 2026, and November 14, 2026.
Mistral launches Vibe Desktop for macOS and Windows, running parallel agent tasks locally or in the cloud
Mistral has released Vibe Desktop, a desktop app for its Vibe agent (formerly Le Chat), for macOS and Windows. It runs multiple tasks in parallel in a local repository or a cloud sandbox, and a Linux version is planned.
Anthropic Adds Experimental Code Mode to Claude Code v2.1.323, Chaining Tool Calls in JavaScript
Anthropic has added an experimental Code Mode to Claude Code v2.1.323. Claude can write JavaScript that chains Bash, Read and Edit calls in one step, and keep talking while long commands run.
Comments
Loading...