我在 12 小时内使用自主 AI 代理构建了一个价值 $150 的 SaaS 计费平台。以下是部署为何成为新瓶颈的原因。

Source: Dev.to

The Challenge: Zero to Production in 12 hours

“Build a production‑ready SaaS billing platform from scratch. No human code. No manual deployment. Go.”

The Result

BillingCore – a functional billing engine with subscriptions, usage‑based metrics, and a dashboard.

The Stack (Chosen Autonomously)

Backend

• Fastify (chosen over Express for lower overhead)
• Prisma
• PostgreSQL
• Redis

Frontend

• React 18
• Vite
• TypeScript

Infrastructure

• Dockerized services deployed to Railway

Auth

• JWT Bearer tokens
• bcrypt

The “Aha!” Moment: Why AI Still Struggles

The Cookie War

• The agent initially used httpOnly cookies for authentication.
• It worked on localhost but failed on Railway where the API and frontend run on different subdomains.
• CORS and browser security policies killed the session, resulting in 401 errors.

How the AI Fixed It

• Without human intervention, the agent analyzed the errors, recognized the cross‑origin limitation, and refactored the entire auth strategy.
• Switched from cookies to Authorization: Bearer tokens across 7 frontend pages and 30+ backend endpoints.

By the Numbers

• Database Schema: 12 tables with complex relations (Subscriptions ↔ Plans ↔ Usage Metrics).
• API: 30+ RESTful endpoints, documented with OpenAPI/Swagger.
• Production Services: 4 (PostgreSQL, Redis, Backend, Frontend).

Lessons Learned

• Autonomous agents can handle end‑to‑end development, but security nuances like cross‑origin authentication still require careful handling.
• Iterative debugging by the agent can replace manual intervention when error patterns are recognizable.

Explore

• Live demo:
• Source code:

What would you ask an autonomous agent to build next? Let’s discuss in the comments.