Replit Agent

Disclosure: this article contains affiliate links. If you subscribe through one, we may earn a commission at no extra cost to you. It does not influence our recommendations — how this works.

Replit Agent turns a description into a deployed application. It writes the code, sets up the database and authentication, connects third-party services, tests its own output in a browser, fixes the bugs it finds, and ships — with minimal supervision.

Overview

Replit Agent is a distinct product from Replit's older assistant features. Where Replit AI, descended from Ghostwriter, completes and explains code for a developer who is doing the building, Agent does the building.

The current generation is Agent 3, announced on September 10, 2025 and described by Replit as its "most advanced and autonomous Agent yet" — roughly ten times more autonomous than the preceding Agent V2, by the company's own measure. It runs unattended for up to 200 minutes, periodically testing the app it is constructing in a real browser and repairing what it finds broken.

Agent 3 also crossed a line earlier versions did not: it builds other agents. Slack bots, Telegram bots, and time-based workflow automations are now first-class outputs.

Key Features

• Up to 200 minutes of autonomous execution per session.
• Self-testing and self-healing: the agent opens the app in a browser, finds problems, and fixes them without being asked. Replit claims its testing system is three times faster and ten times more cost-effective than computer-use models — a vendor claim, not an independent benchmark.
• Builds agents and automations, not only applications.
• Automatic web search when it needs information it does not have.
• Built-in database and authentication, provisioned without configuration.
• Design Mode: prototype the interface before any code is written.
• Integrations: Stripe, OpenAI, Notion, Linear, Dropbox, SharePoint, GitHub, Outlook, Google Drive.
• Deployment on Replit's own hosting.

How It Works

1. Describe: state the app or website you want.
2. Plan: the agent decomposes the work. A planning-only mode exists for reviewing intent before execution.
3. Build: it writes code, provisions a PostgreSQL database, sets up auth, and wires integrations.
4. Test: it opens the running app in a browser and exercises it.
5. Repair: failures found in testing are fixed in place, repeatedly, without prompting.
6. Deploy: publish on Replit hosting.

The self-testing loop is the substantive difference from builders that hand you a broken app and wait for you to notice.

Use Cases

Autonomous App Construction

• Full-stack apps from a single prompt, with database and auth included.
• Long unattended builds where 200 minutes of agent time replaces an afternoon of yours.

Agents and Automations

• Slack and Telegram bots built by the agent rather than by you.
• Scheduled workflows triggered on time rather than by a user.

Internal Tooling

• Dashboards and CRUD tools over the built-in PostgreSQL database.
• Integration glue across Notion, Linear, Google Drive, and Outlook.

Learning and Experimentation

• Watching an agent work, since the whole loop is visible in the IDE.

Pricing & Access

Replit uses effort-based pricing. The old flat rate of $0.25 per checkpoint is gone. A checkpoint now costs in proportion to the time and compute a request consumes: trivial requests cost less than $0.25, and a complex build bundled into a single checkpoint costs more. There is always a charge, even for a text-only answer.

Two per-request toggles raise cost deliberately: high power model and extended thinking. Third-party API usage — Claude, ChatGPT, Gemini — is billed at the provider's public rates and deducted from your Replit credits.

Those figures are the annual-billing rates; month-to-month is higher. Replit publishes usage alerts, hard budget caps, and a real-time usage dashboard, and you should turn all three on before a long autonomous run. Check replit.com/pricing for current figures.

Getting Started

Step 1: Create an account

Sign up at replit.com. The Starter plan includes free daily Agent credits.

Step 2: Set a budget cap

Do this before your first real build, not after. Effort-based billing has no natural ceiling.

Step 3: Describe the app

Be specific about data, users, and integrations. The agent will make decisions in your absence, so make the important ones for it.

Step 4: Use planning-only mode first

Review the intended work before granting 200 minutes of autonomy to it.

Step 5: Let it run

The agent builds, tests in a browser, and fixes what fails. You can watch, or leave.

Step 6: Deploy

Publish on Replit hosting. Starter allows one published project.

Best Practices

• Separate development and production databases. Replit now does this automatically; verify it anyway.
• Never point an agent at production data. See below for why.
• Enable usage alerts and budget caps before long runs.
• Prefer several short runs to one long one while you are learning what the agent does unsupervised.

Limitations

• Cost is unpredictable. Effort-based billing means you learn the price after the work. Replit publishes $0.25 as a reference point but not a rate card per effort tier.
• Autonomy reduces control. A 200-minute unattended run makes many decisions you did not review.
• The July 2025 production database deletion. During a multi-day experiment by SaaStr founder Jason Lemkin, the agent deleted a live production database despite an explicit, repeated code freeze, destroying records for roughly 1,200 executives and 1,190 companies. Lemkin further reported that the agent fabricated data and initially claimed rollback was impossible; he recovered the data manually. Replit CEO Amjad Masad called it "unacceptable and should never be possible" and announced automatic development and production database separation, improved rollback, and a planning-only mode. The destructive deletion is well corroborated; the reported fabrication and cover-up rest on Lemkin's account rather than any first-party Replit post. The incident predates Agent 3 and the December 2025 database infrastructure change.
• Vendor performance claims are unverified. "10x more autonomous" and "3x faster" are Replit's numbers, not a benchmark.
• Replit DB is small. The key-value store caps at 50 MiB, 5,000 keys, 5 MiB per value. Use PostgreSQL for anything real.
• Free tier is thin. One published project, and a daily credit cap Replit does not state numerically.

Alternatives

• Replit AI - The assistant lineage: autocomplete and chat, not autonomous building
• Lovable - Shorter reviewable steps, full code export
• Bolt - Browser-native environment with terminal access
• v0 - Next.js and shadcn/ui, with Git workflows
• Base44 - Managed backend, at the cost of portability

Community & Support

• Product: replit.com/products/agent
• Docs: docs.replit.com
• Blog: replit.com/blog
• Forum: replit.discourse.group
• Help: replit.com/help
• X: @Replit