Replit says a swarm of AI agents built a working application monitoring platform in two days for $3,000. That is the headline of Replit's launch of Missions on November 17, 2026, and the numbers come from Replit's own tests. This post explains what Missions is, lists each claim, and says what could be checked.
The primary sources are Replit's post Introducing Replit Missions (published November 17 and updated November 19, 2026) and the Missions product page.
TL;DR: claims and checks
| Claim (Replit) | Detail | Status |
|---|---|---|
| DataCat built in 2 days | Within a 3,000 dollar capped budget | Replit statement; no independent rebuild |
| Planted DB bug alert | 38 seconds against a 60-second goal | Replit test |
| 3-hour load test | 26M records, 4.4M log entries, zero loss | Replit test |
| Checks passed | 51 of 51 data and alert checks | Replit test |
| Versus a Claude Code build | 5x trace intake, about 2x metric intake; Claude Code build faster on logs | One run per setup, Replit scoring |
| NumPy | 4x solo, 17x with 16 agents, 19x with 32 agents | Replit benchmark on 13 tasks |
| Price and plans | Not published | Gap |
What Missions is
Replit describes Missions as a way to "take on the complex work they haven't had time or capacity for." You define the outcome. Replit coordinates a swarm of agents to plan, build, test and iterate toward it, within the budget and boundaries you set. Examples in the post: winning your first 10 paying enterprise customers, improving product performance, replacing legacy software.
The control model is the interesting part:
- Mission agreement. Before work starts you agree on goal, scope, success criteria, budget and which actions need your approval.
- Budget cap. The post gives an example of capping spend at $1,000 and requiring approval before the agent contacts prospects, shares a proposal, changes customer-facing materials or makes product commitments. The product page says the mission pauses at the limit and asks before raising it.
- Independent judge. A separate AI judge checks results against the agreed criteria. A mission is complete when the criteria are met. If the budget runs out first, work stops with the outcome unresolved.
- Updates. You choose what you hear about (blockers, progress, status) and how often: immediately, daily or weekly. Cloud agents keep running if you close your laptop.
- Output. The product page says changes arrive as pull requests and reviewable files, and anything outside the boundaries waits for a person.
Missions differs from Routines, which run on a schedule. A mission works toward an outcome, such as improving signup conversion. To start, click "New" in the sidebar, choose "Mission", and describe the goal.
Claim 1: DataCat in two days for $3,000
Replit wanted to bring application monitoring in-house. It gave Missions the goal of building DataCat. According to Replit, in two days and within a $3,000 budget, a swarm built dashboards, alerts and application tracing, then tested the platform against measurable goals. Missions also built DataCat's own agent, which collects data from each server, and its tracing libraries.
Replit's reported results:
- A deliberately introduced database bug raised an alert in 38 seconds, beating a 60-second target, and was traced to the exact query and logs.
- A three-hour load test with sample apps stored 26 million activity records and 4.4 million log entries with no data lost, while searches and alerts kept running.
- All 36 data-correctness checks and 15 alert-reliability checks passed (51 of 51).
- In matched Node.js tests, the app handled 39% more requests with 51% less monitoring-related CPU overhead.
- DataCat's agent used 82 to 108 MB of memory under load, against 319 MB for the original platform's agent, and made logs searchable in 0.5 seconds against 4.2 seconds (median).
Replit adds a caution in its own notes: these controlled tests "do not establish performance against the original platform's hosted service or on Replit's production systems." The agent comparison used the original agent version 7.56.2. The post names the reference server it compared against only as "Radar."
Claim 2: DataCat versus a Claude Code build
Replit compared DataCat with a monitoring platform built from the same initial prompt using /goal in Claude Code. Its table says:
| Metric | DataCat | Claude Code build |
|---|---|---|
| Data types accepted from the original agent | Logs, metrics, host data, traces | Logs only |
| API compatibility checks | 51 of 56 per storage engine | 45 of 56 |
| Trace intake | 49,000 per second (Postgres) | 9,800 per second |
| Metric intake | 405,100 points per second | 224,600 |
| Storage engines | 2 (Postgres, ClickHouse) | 1 (SQLite) |
Replit also says the Claude Code build "covered more product areas and took in logs faster." That is a useful honest line. Still, note the method: one run per setup on identical cloud machines, scored only on shared product areas. A single run of an agent is noisy, and the scoring was done by the party that sells one of the two tools. Treat the 5x as a Replit data point, not a head-to-head benchmark. For other recent cases where a vendor's own table needs checking, see our posts on agentic QA claims and Rust rewrites checked against recorded parity.
Claim 3: more agents, faster NumPy
Replit tested swarms on 13 representative NumPy benchmark tasks. Agents explored alternative implementations, optimized memory access and validated correctness automatically.
With Auto (Max), one agent reached speedups up to 4x, a 16-agent swarm 17x and a 32-agent swarm 19x. With Auto (Power): 2x, 7x and 9x. Replit calls this "the striking result": exploring more approaches in parallel found much faster code.
What is not stated: which 13 tasks, how speedup was aggregated, the cost per swarm size, or whether the results hold on other libraries. "Up to" figures hide the spread across tasks. The shape of the result matches research we covered on agents that improve their own harnesses: more parallel search finds more, at higher cost.
Other claims in the post
- E-Commerce Bench. Replit says an agent starting with 100,000 yen reached 4.82 million yen over a simulated year in one run, about 48 times the start and, in its words, more than three times the best on the public leaderboard. It is a single run on a simulation.
- Replit Slides. The post includes a figure of agent swarms working on improving Replit Slides, but the text gives no numbers for it.
- Self-Driving Company. Replit frames Missions as a step toward a company where people set direction and stay accountable while agents execute.
What is not published
- Pricing and plans. The product page says you set a maximum budget and that work counts toward it. It does not say which plans include Missions or any per-mission rate.
- Model mix. The post refers to Auto (Max) and Auto (Power) modes, not to which models run underneath.
- Independent results. No third party has reproduced DataCat or the NumPy numbers in the sources we reviewed.
- Enterprise-sales missions. The "win 10 paying customers" example is a scenario. The post shows no customer result.
- Failure rates. There is no count of missions that ran out of budget before meeting criteria.
Why the control design matters more than the headline
The cost story is easy to over-read. A $3,000 cap is a ceiling the user sets, not the price of any outcome. What stands out is the structure: a contract before work, a cap that pauses the run, approval gates for outward actions, and a judge that is separate from the workers. That is the same shape enterprises ask for elsewhere. Compare BMW's per-step autonomy levels and CI/CD gates and AWS session budgets for agent purchases. In all three, spending and permission limits sit outside the agent's own reasoning.
How to check these claims yourself
- Start with a goal you can score. Replit's own advice is to name the work and how you will judge it, such as rebuilding a workflow against a feature checklist.
- Set a low cap first. Use the product page's pattern: cap, approval before publishing, and a limit-increase review.
- Re-run the comparison. If you test against another agent, run each setup several times and score on the same checks. Replit used one run per setup.
- Read the DataCat repo. Replit says a self-hostable version is open-sourced; read the code and license before relying on it.
- Inspect the PRs. Because changes arrive as pull requests, review them like any contributor's work.
What to do now
Builders: Try one mission with a small cap on a task with an existing test suite, such as an API latency target or a Rust rewrite with parity checks.
Founders: Do not budget from the $3,000 figure. Ask Replit for plan inclusion and per-mission pricing, then run a capped pilot.
Marketers: Treat the sales and SEO mission templates as drafts needing approval gates. Replit's own settings require approval before outreach is sent.
Sources and related reading
- Replit, Introducing Replit Missions (November 17, 2026, updated November 19)
- Replit, Missions product page
- Rust anything: incremental rewrites with recorded parity
- LLM agents that self-improve without a prescribed loop
- BMW Group agent governance
