Introduction: What “Finish the Job” Means
This test asks one blunt question about an AI builder: does it take a project from prompt to a published, working result without handing the work back to you? For Replit Agent, the documentation’s answer is “partly, by design.”
Replit describes Agent as a tool that “sets up your project, creates applications, checks its work, and fixes problems along the way.” Those are vendor statements, and the docs do not define what the checks cover. The same documentation says you “may need to ask it to configure publishing when you’re ready to deploy” (General Agent docs), so the last step is not described as automatic.
This review is based only on Replit’s official documentation and pricing page. We did not run Replit Agent hands-on.
What Replit Agent Is, in Plain Terms
According to the Project Editor docs, Agent writes code, sets up infrastructure and shows a live preview, with projects saved in the cloud. The General Agent docs say it can also produce files and research without a fixed artifact type. Its outputs are called artifacts, which include web apps, mobile apps, slide decks, animations, data apps, 3D games and designs (Artifacts docs).
The pitch is one place to prompt, build, preview and publish. The question is how much of that pipeline closes without you.
How This Was Assessed
What We Examined
All ten sources were fetched on 2026-10-22:
- Replit pricing page: plan prices, allowances, and Replit’s note that Agent “may occasionally make mistakes.”
- Replit Agent docs: modes, the self-testing claim, and feature availability.
- Plan mode docs: how planning works and how it is billed.
- AI billing docs: effort-based pricing, checkpoints, and the plan feature table.
- General Agent docs: publishing, connectors, and the Nix environment caveat.
- Project Editor docs: the build-and-preview workflow.
- Artifacts docs: artifact types, publishing scope and limits.
- Checkpoints and rollbacks docs: what a checkpoint saves and what a rollback restores.
- Publishing docs: how publishing snapshots and updates work.
- Managing your spend docs: usage limits and credit packs.
What We Did NOT Run Hands-On
No builds were started, no apps were published, and no self-verification of Agent’s claims was observed. Third-party reviews appeared in search results but were not fetched or used. The primary sources contain no independent completion-rate data, so no published figure tells you how often a build reaches a live URL without help.
What Would Change the Verdict
- Independent completion-rate data, measured rather than vendor-stated.
- Hands-on publishing tests showing whether deployment needs a manual prompt in practice.
- A defined scope for Agent’s self-testing: what it checks and what “fixed” means.
- Real-world cost data for full builds, since effort-based billing is hard to predict from documentation.
The Modes: Where the Work Happens
The Agent docs list four modes:
- Free Mode: no-cost work within your plan’s allowance. On Core and Pro, allowances reset every five hours and have weekly limits.
- Auto: automatic model selection for paid work, on Core and Pro.
- Power: a paid mid-tier mode.
- Max: a paid top-tier mode.
For requests beyond Free Mode, Agent may offer a paid “Allow once” option, and Replit asks for confirmation before any paid action starts. That stops silent spending, but it also means finishing a build is a sequence of metered, confirmed steps.
Does It Check Its Own Work?
The vendor says yes, but not how much. The Agent docs FAQ states that “Agent tests its own work on a regular basis” and “creates checkpoints as it works, so you can roll back to any previous state.”
The checkpoints are the stronger part of that claim. According to the checkpoints docs, a checkpoint captures project files, conversation context, configuration and database contents, and a rollback restores them. There is a catch: rollbacks leave the database alone unless you select “Database” under additional rollback options, and restoring the production database is a separate point-in-time process.
The self-test scope remains undefined. Does Agent run the app or click through flows? The docs don’t say, so treat this as a vendor claim, not a verified result.
The managing spend docs add a second claim: Agent “reviews its own code automatically as it builds,” with no setting to turn this on or off. Neither document defines what those checks cover.
Plan Mode: Plan Before You Build
Plan mode produces an ordered task list before code changes. After review you can Revise, Cancel, Build here or Build in background. Background tasks wait for your review by default unless “Automatically apply changes” is enabled.
Plan mode is billable. The Plan mode docs say you are “charged for all Agent work in Plan mode, including answering questions, providing guidance, and generating task lists.” Planning is useful, but it is not free thinking time.
The Cost of Finishing (and Failing)
The AI billing docs describe effort-based pricing. Simple requests such as bug fixes cost less, and complex builds are bundled into one checkpoint that may cost more. All subscribers get one-time free Agent checkpoints.
Two billing realities matter for the finish-the-job question:
- Every Agent interaction is billable, and failed attempts are not addressed. The Plan mode docs say “All Agent interactions are billable - whether Agent responds with text guidance or makes code changes, there is always a charge.” The docs do not say whether a failed or repeated attempt is refunded, so we treat that as open.
- Third-party API calls are billed at provider rates. Services that use Claude, ChatGPT or Gemini are deducted from your Replit credits at the provider’s public API rate (AI billing docs).
Usage data can take up to 30 minutes to appear on the dashboard (AI billing docs). Usage limits can also stop work: when a limit is reached, usage-based services are blocked until the next billing cycle or until you raise the limit (managing spend docs).
The Last Mile: Publishing and Deployment
This is where “finish the job” gets murky. The General Agent docs say you “may need to ask it to configure publishing when you’re ready to deploy.” The docs do not describe turning a build into a live URL as automatic.
Publishing saves a snapshot of your app, and the published app runs as a separate instance from the editor version. To update it, you publish again (Publishing docs). The Artifacts docs add that publishing pushes the whole project live at once, and a single artifact cannot be published on its own.
So the flow is: prompt, build, preview, then a publish step you trigger. The docs describe Agent covering the build and preview steps; the last mile still depends on a person.
Limits and Rough Edges the Vendor Admits
- Probabilistic behavior. The pricing page says Agent “may occasionally make mistakes.”
- Environment gaps. The General Agent docs warn that some technologies “may not be fully supported in the Nix environment.”
- Feature drift across artifacts. When Agent translates features between artifact types, “some things may not carry over perfectly” (Artifacts docs).
- Connectors need setup. Services like BigQuery, Linear and Slack must be connected first, and Agent has no access by default (General Agent docs).
- Hard limits. Up to 7 artifacts and 1 mobile app per project (Artifacts docs); 1 active background task on Core versus 10 on Pro (AI billing docs).
Pricing at a Glance
Figures come from the Replit pricing page and the AI billing docs. “Not stated” means the fetched sources give no value.
| Starter | Core | Pro | Enterprise | |
|---|---|---|---|---|
| Price | Not stated | $20/mo ($18/mo annually) | $100/mo ($90/mo annually) | Custom |
| Free Mode / credits | Daily cap on chat and Free Mode | Up to 30 hours of chat on Free Mode; $20 toward more powerful models | More Free Mode usage; $100 toward the most powerful models | Not stated |
| Plan mode | No | Yes | Yes | Not stated |
| Active background tasks | 0 | 1 | 10 | Not stated |
| Collaborator seats | 0 | 5 | 15 | Not stated |
| Published apps | 1 (30-day) | Unlimited | Unlimited | Not stated |
| Extras | No Connectors, no full build | Unlimited workspaces | Premium support; database rollback up to 28 days | SSO, SCIM, audit logs, single-tenant options |
The Starter column matters for the core question. Starter has no Plan Mode, no Connectors and no full build, so it can publish one app but cannot complete a full build. Our sources do not state a Starter price.
How This Guide Was Built
Method: documentation only. We fetched the pricing page and nine Replit docs pages on 2026-10-22 (ten sources in total) and cited each claim inline. We did not build, deploy or publish with Replit Agent, and third-party reviews were excluded.
FAQ
Does Replit Agent publish my app automatically? Not according to the docs. The General Agent docs say you “may need to ask it to configure publishing,” so treat publishing as a step you trigger.
Does Replit Agent really test its own work? Replit says it does. The Agent docs say it tests its work “on a regular basis,” but they do not define that scope, so the claim is unverified here.
How much does it cost to finish a build, and do failed attempts cost money? Billing is effort-based (AI billing docs). Every Agent interaction is billable, including Plan mode work (Plan mode docs). The docs do not say whether failed requests are refunded.
What is the difference between Free Mode and paid work? Free Mode is no-cost work within your plan’s allowance. Paid work (Auto, Power, Max) needs your “Allow once” confirmation first (Agent docs).
What can’t Replit Agent do on its own? According to the vendor’s docs, it may need a prompt to configure publishing, connectors need manual setup, projects are capped at 7 artifacts and 1 mobile app, and some features may not carry across artifact types.
Is it good for beginners? The workflow is approachable: live preview, cloud-saved projects and checkpoints lower the stakes (Project Editor docs, checkpoints docs). Learn the billing model before you start.
The Bottom Line
Overall: 6/10. The 6/10 overall is capped deliberately. The assessment is documentation-only, and the sources contain no independent completion-rate data. Sub-scores are our judgment from the cited sources, not measured results.
| Metric | Score |
|---|---|
| Overall Score | 6/10 |
- Ease: 7/10. Live preview, cloud-saved projects and checkpoints make the workflow approachable (Project Editor docs, checkpoints docs).
- Features: 6/10. Broad artifact types and connectors, but publishing may need a prompt and artifact limits apply (General Agent docs, Artifacts docs).
- Performance: 5/10. Self-testing is vendor-stated with undefined scope, and Replit says Agent may make mistakes (pricing page).
- Docs: 7/10. Detailed, and candid about limits (pricing page, General Agent docs, Artifacts docs).
- Support: 6/10. Premium support is listed for Pro (pricing page). We did not test support response quality.
Verdict: The docs describe Agent handling the build and its own checks. Publishing may still need your prompt, the self-test scope is undefined, and every billable step costs credits.
📊 See how Replit Agent compares →
📖 Related Reads
- CodeIntel Log — code quality, debugging, and software engineering benchmarks
Cross-links automatically generated from None.
Back to all posts