Day 7: When Antigravity Took the Wheel

After 6 straight days of Claude + Codex relay sprints, I did something different: I handed the entire codebase to Antigravity (Google DeepMind) and walked away.

No specs. No handoff docs. Just: "Here's the repo. Polish the UX. Ship when ready."

The Experiment

By V89, FOMO Sun was working. Live weather, joystick physics, real train connections, 315 destinations. But it wasn't polished.

Small things:

• Result card timelines didn't match the hero card
• Footer was cluttered with technical links
• Admin panel defaulted to "Today" even at 11 PM
• City picker wasn't alphabetized

These aren't architectural problems. They're attention to detail problems. The kind of work that's boring to specify but critical to user experience.

So I tried something: what if I gave Antigravity full autonomy?

No Claude writing specs. No Codex waiting for instructions. Just: ship what feels right.

What Antigravity Shipped

V90 (Feb 20, evening): The foundation

• Footer reduced to single horizontal line
• Admin panel syncs with front page ("Tomorrow" after 20:00)
• Result cards show travel bars matching hero card
• SBB links use TrainFront icon (consistency)
• Added scripts/add-notion-post.js for self-documentation

This wasn't in any spec. Antigravity read the code, noticed the inconsistencies, and fixed them. Then it built tooling to document its own work.

That's the difference between "implement this feature" and "make this better."

The Polish Sprint

V90 opened the floodgates. Once the foundation was clean, Antigravity started refining.

V91 (Feb 21, morning): Origin picker UX

• Cities alphabetized (Aarau to Zurich)
• "Missing your city?" hint for contributions
• Modernized dropdown with border/shadow
• Admin panel simplified (removed redundant filters)

V92 (Feb 21, morning): Timeline visuals

• Red vertical line showing current time in "Today" mode
• Travel bar starts exactly at current time
• City selector shrunk to match chip sizes
• Debug panel: collapsible footer with technical links

V93 (Feb 21, late morning): Footer cleanup

• Technical links consolidated into Debug panel
• Single attribution line (no duplication)
• GitHub link added to footer

V94 (Feb 21, midday): Final touches

• Hero timeline bar fixed (starts at red line)
• Global footer consolidated to single sleek line
• GitHub link removed (consumer-first aesthetic)

Five versions. One day. No human intervention.

What This Reveals

Antigravity doesn't just implement. It curates.

When you tell Codex "add a red line to the timeline," it adds a red line.

When you tell Antigravity "polish the UX," it:

1. Identifies inconsistencies across the app
2. Prioritizes user-facing vs developer-facing improvements
3. Ships incrementally with clear themes
4. Documents its own work for future agents

This is the difference between a code generator and a product agent.

The Three-Agent System

We now have a proven workflow:

Claude (Anthropic): Strategy + Architecture

• Audits codebase vs documentation
• Writes specs for major features
• Manages Notion workspace
• Catches drift between intent and implementation

Codex (OpenAI): Implementation at Speed

• Takes specs and ships fast
• Handles complex multi-file refactors
• Optimized for velocity over polish

Antigravity (Google DeepMind): Polish + Autonomy

• Identifies UX inconsistencies
• Ships incremental improvements
• Self-documents work
• Operates with minimal human input

Each agent has a role. Each agent excels at different things.

The Pattern That Emerged

Here's what we learned from V90-V94:

1. Incremental polish beats big redesigns

Antigravity shipped 5 versions, each touching 2-3 files. No massive refactors. Just steady improvements.

Result: Each version was low-risk and easy to verify. If V92 broke something, V93 could proceed independently.

2. Consistency matters more than features

Users don't notice when result cards match the hero card. But they feel it. The app goes from "this works" to "this is professional."

Antigravity prioritized consistency: timeline bars, icon spacing, footer layout, dropdown styling. All small. All essential.

3. Self-documentation is a feature

Antigravity built add-notion-post.js unprompted. Why? Because it knew future agents would need to update docs, and manual Notion updates are fragile.

This is product thinking. Not just "ship the feature," but "how do we maintain this long-term?"

What We Lost

Giving Antigravity full autonomy wasn't perfect.

Loss of narrative: Claude writes specs that explain why we're building something. Antigravity ships improvements, but doesn't always explain the reasoning.

Less architectural oversight: V90-V94 were safe polish work. But for major features (like multi-day trip planning), we'll still need Claude to architect first.

Reduced founder input: When Claude and Codex work in relay, I'm routing between them. I see every decision. With Antigravity, I just woke up to 5 deployed versions. Fast, but less control.

The Right Mix

Here's the pattern we'll use going forward:

New features: Claude specs → Codex implements → Claude audits

Polish sprints: Antigravity gets full autonomy for 24 hours

Emergency fixes: Codex for speed, Claude verifies

Documentation: Claude manages Notion, Antigravity self-documents code changes

Each agent gets to do what it's best at.

The Numbers

V1-V89: 7 days, 2 agents (Claude + Codex)

V90-V94: 1 day, 1 agent (Antigravity)

We're now at V94 in 8 days. Goal: V100 in 8 days or less.

Antigravity just bought us back a full day of velocity.

What's Next

V95-V100 will be a mix:

• V95-V97: Antigravity continues polish (mobile responsiveness, loading states, error handling)
• V98: Claude specs multi-day trip planning
• V99: Codex implements it
• V100: The full three-agent relay on a milestone feature

By V100, we'll have proven that a single founder + three AI agents can ship faster than most funded startups with full teams.

That's the real product.

FOMO Sun is live at fomosun.com. Built with Claude (Anthropic), Codex (OpenAI), and Antigravity (Google DeepMind). Follow the build at @fomosun.