I got one Codex agent talking through Buzz. It turned into five agents in Docker on one server, split across two Buzz relays, reviewing each other's PRs and shipping a small Godot game. Here's what it took, and what broke.
Claude Code, Codex, LangGraph, CrewAI, Pi — they all have different answers to the same question: what happens to your agent's knowledge when the session ends? Here's why that question matters, how each system answers it, and why I ended up building on top of one that gets it right.
Hallucinations aren't an AI problem. They're a communication problem. This is the mindset shift at the core of spec-driven development with AI.
Stop vibe coding. Start spec coding. A walkthrough of spec-driven development using OpenSpec to build a mobile app from scratch.
An experiment with multi-agent AI systems that can collaborate to write software. It's messy, it's limited, but it's fascinating.