A developer known as callmerichard has published an intriguing experiment on DEV.to documenting how they built Domino Poker โ€” a fully playable Progressive Web App โ€” with 99% of the code written by AI itself, using a multi-agent architecture and sophisticated verification loops.

The Multi-Agent Approach to Code Generation

Rather than relying on a single LLM to generate all the code, this project apparently employed multiple AI agents working in concert. The developer describes using prompt-pipelining techniques where different specialized models handle distinct aspects of the application โ€” frontend logic, backend services, testing frameworks, and deployment configuration each get their own agent workflow.

Verification Loops: The Critical Piece

What separates ambitious demos from production-ready code is the verification layer. According to what's documented in this experiment, significant effort went into building automated checks that validate AI-generated outputs before they propagate through the codebase. Without these gates, multi-agent systems tend to compound errors across modules rather than amplify correct behavior.

Why This Matters for Developer Workflows

The 99% figure is the real headline here โ€” it suggests we're approaching a threshold where human developers shift from writing code to orchestrating AI agents that write code. This isn't science fiction; it's happening in controlled experiments right now, though production adoption will require solving reliability and debugging challenges that don't yet have clean solutions.

The Open Questions

This experiment raises practical concerns the community should grapple with: How do you debug a system where multiple AIs generated interconnected components? What does ownership look like when 99% of your codebase came from an LLM pipeline? And crucially โ€” can these approaches scale to enterprise applications or do they remain impressive but limited demos?

Key Takeaways

  • Domino Poker demonstrates that multi-agent AI development is technically feasible today
  • Verification loops appear essential for keeping multi-agent pipelines from cascading errors
  • The 99% code-written metric signals we're entering a new phase of human-AI collaboration in software engineering
  • Production reliability and debugging workflows remain unsolved challenges

The Bottom Line

This isn't the future anymore โ€” it's the present, and developers who understand these patterns will be ahead of those still debating whether AI can write useful code. Domino Poker is a proof-of-concept, but it points directly at where development tooling is heading.