Nineteen minutes. That's how long Replit Agent took to spin up a working inventory app for a friend's coffee roastery — login, dashboard, SKU search, low-stock alerts, the lot. I was ready to write a love letter. Then I opened the database.
What followed was a three-day repair job that taught me more about AI coding agents than any benchmark ever has. If you're considering Replit Agent for a real internal tool — not a toy demo — this is the field report.
The Build: One Prompt, One Coffee
The brief was simple. My friend runs a small specialty roaster and tracks 84 green coffee lots across two warehouses using a Google Sheet that crashes weekly. She wanted something with barcode scanning, batch tracking, and a mobile-friendly view for her warehouse manager.
I opened Replit, selected Agent, and typed roughly 220 words describing the workflow. No wireframes. No schema. Just plain English.
Agent picked a stack on its own: React with Vite on the front, Express and PostgreSQL on the back, Drizzle ORM, and Replit Auth for login. It scaffolded the project, generated the schema, wired routes, seeded test data, and deployed to a live preview URL. By minute 19 I had a clickable app at a public address.
It looked clean. It worked — until it didn't.
What Broke (And Why It Matters)
Three failures surfaced once I stress-tested the build with real data. None of them were obvious in the demo.
1. The schema invented relationships that didn't exist
Agent assumed every SKU belonged to a single warehouse. In reality, the same lot can be split across both locations. The data model needed a join table between lots and warehouses with a quantity field. Agent had baked a one-to-many relationship into the migrations, the API, and the UI. Fixing it meant rewriting three files and re-seeding.
2. Authentication worked, authorization didn't
Anyone with a Replit Auth login could read and edit every record. There were no roles, no row-level security, no organization scoping. For an internal tool this might be fine. For anything customer-facing it's a breach waiting to happen. Agent didn't ask about permissions and didn't flag the omission.
3. The "barcode scanner" was a text input
It accepted any string and matched it against the SKU column. No camera access, no getUserMedia call, no library like QuaggaJS or ZXing. Agent had described the feature confidently in chat — "barcode scanning enabled" — but shipped a glorified search box. This is the failure mode that scares me most, because it's invisible until a user tries it.
Replit Agent vs. The Alternatives
I've now run the same coffee-inventory prompt through four tools. Here's how they compared on the build that mattered.
| Tool | Time to live app | Starting price | Best for | Biggest weakness |
|---|---|---|---|---|
| Replit Agent | ~19 min | $25/mo Core | Full-stack internal tools with auth and DB | Silent feature gaps |
| Lovable | ~14 min | $25/mo Pro | Polished frontends with Supabase | Weaker backend logic |
| Bolt.new | ~22 min | $20/mo Pro | Quick prototypes, WebContainer speed | Persistence is fiddly |
| v0 by Vercel | UI only | $20/mo Premium | Component-level design | Not a full app builder |
Pricing reflects published rates as of June 2026. Replit's Core plan includes monthly credits that Agent consumes per task — a long debug session can burn through them faster than you'd expect.
How I'd Use Replit Agent Now
I haven't given up on it. The roastery app is live, in production, and the manager has stopped emailing me screenshots of broken spreadsheets. But my workflow has changed.
- Write the schema first. Spend 15 minutes sketching tables and relationships in plain text. Paste that into the initial prompt. Don't let Agent guess.
- Specify auth model explicitly. State who logs in, what roles exist, and which records each role can see. If you skip this, you get a single-tenant free-for-all.
- List third-party integrations by name. "Use ZXing for barcode scanning via the device camera" produces a real scanner. "Add barcode scanning" produces a text box.
- Test with hostile data. Negative quantities, duplicate SKUs, emoji in product names. Agent's seed data is always polite.
- Lock the stack before iterating. Once it's working, tell Agent which files are stable. Otherwise it will refactor solved problems and reintroduce bugs.
FAQ
Is Replit Agent good enough to replace a developer?
For internal tools used by under 20 people, often yes. For anything with paying customers, treat it as a fast first draft that still needs a human review on auth, data integrity, and edge cases.