AI / Wanderings 2025
Vibecoding
Google's Antigravity shifts software work from writing every line to describing, reviewing and testing what agents build. This is the practical workflow.

A growing part of software development now begins with a description rather than a blank code file. Agents implement the first version; the human reviews the architecture, behaviour and trade-offs.
Google Antigravity is one tool for this orchestration workflow.
From Code-Writing to Orchestration
Traditional dev workflow: you sketch the feature, write the code, debug the code, refactor the code, test the code, ship the code.
Vibecoding workflow: you describe what you want, the agents write the code, you review the code, you test the behaviour, you ship.
Your main job moves from implementing every line to describing the solution, setting constraints and checking what the agents produce.
That shifts where the judgement sits.
Much of traditional coding consists of boilerplate, familiar patterns and searching through Stack Overflow. Agents excel at that work. And when agents handle that, you're left with the 30% requiring human judgement: "is this the right approach?" "does this match the intended behaviour?" "does this feel right?"
The stack that works
You don't need to use Google Antigravity. But if you're building web products, here's the practical foundation:
Frontend: Next.js + Tailwind
- Next.js gives you routing, server components, deployment to Vercel in one click
- Tailwind gives you styling without thinking about CSS architecture
- Together they're the fastest path from "idea" to "something people can use"
Database: Supabase or Cloud SQL
- Supabase if you want managed PostgreSQL with auth baked in
- Cloud SQL if you're already in Google Cloud
- Either one: you get a real database without managing infrastructure
Payments: Stripe Checkout
- Antigravity has built-in workflows for Stripe
- You drag and drop "collect payment," it generates the integration
- This saves weeks of integration work
Deployment: Google Cloud Run
- One-click deploy from your local environment
- Scales to zero cost when you're not using it
- Antigravity has native Cloud Run workflows
Mobile: Capacitor
- Convert your web app to iOS/Android without rewriting
- Takes a Next.js app and generates native wrappers
- You're not maintaining two codebases
The Antigravity Workflow
Start in Planning Mode. Master prompt approach.
You describe what you want. Detailed but conversational. "I want an e-commerce app where users can browse products, add to cart, check out with Stripe, and see their order history."
Antigravity breaks this into:
- Data models (Product, Order, OrderItem, User)
- API endpoints
- Frontend components
- Workflow integrations
You review each piece. You can say "change the Product model to include inventory_location" and it regenerates everything downstream.
Then you connect to the MCP Store, managed connection providers for AlloyDB, Cloud SQL, etc. One click. You're connected to a real database.
Drop in a Stripe Checkout workflow. Drop in a Cloud Run deploy workflow. You've got a complete pipeline.
Why agent-generated code can work
AI-generated code can be bad. But the code Antigravity generates is better than average junior developer code. It's boring, clean, follows conventions, includes error handling. The generated code favours competence over creativity.
You still review it. When something is wrong, you fix it or regenerate the relevant part. The feedback loop stays tight.
Compare that to traditional dev: you write code, you test it, you find bugs three weeks later in production. The feedback loop is slow.
The iteration advantage
The larger speed advantage comes from faster iteration, beyond the time saved writing code.
You want to try a different UI layout? Five minutes. You want to change the database schema? Regenerate the models and migrations. You want to add a new workflow? Describe it, review it, ship it.
This lets you explore the product space instead of committing to an initial design and defending it. You can build three different approaches and see which one feels right. That's how you discover what users want instead of what you thought they wanted.
What to unlearn
This requires unlearning something: the idea that code quality is measured by cleverness.
Good code is boring. It works. It's maintainable. It handles edge cases. Most importantly, it solves the right problem.
Antigravity excels at boring code and struggles with novel algorithmic problems. Many products mainly need data to move through systems in sensible ways, which suits agents well.
Getting Started
If you're building something today:
- Describe it to Antigravity in detail (one paragraph per feature)
- Review the generated schema and endpoints
- Build the frontend component by component
- Connect to Cloud SQL or Supabase
- Add payment flows
- Deploy to Cloud Run
- Wrap with Capacitor if you need mobile
Total time: 2–3 weeks for an MVP that would take months with traditional development. Faster iteration can also produce a better product because you test more approaches before settling.
What this changes
This means traditional software engineering, the years of learning frameworks, optimisation, debugging, has less moat.
But it also means you can build what used to require a team of five developers. You and an AI, moving fast, iterating, discovering.
That matters when your aim is to build a product people love rather than become an expert in the latest .js framework.
Start with one bounded feature. Write the behaviour and constraints in plain language, inspect the generated schema and code, then test the result as if a junior developer had submitted it for review.