Steve Yegge's GasTown and Dan Lorenc's multiclaude took different paths but arrived at identical architectural primitives for multi-agent coding systems. Both use git worktrees for isolation, tmux for observability, and CI pipelines as progress ratchets.

Running multiple AI agents on shared codebases creates distributed systems challenges: crashes, conflicts, and state loss. These projects solve persistence through git-backed hooks or filesystem storage, ensuring work survives failures. Their convergence suggests foundational requirements are emerging.

One key insight is scoped autonomy with external persistence. Agents get bounded freedom to operate, with CI validating progress rather than real-time coordination. GasTown demands active human oversight while multiclaude enables asynchronous collaboration like an MMORPG.

These patterns point toward infrastructure-first agent design. Expect tooling to evolve around isolation, observability, and failure-resistant state management. Teams building AI orchestration should watch these primitives closely—they'll outlast current pricing and prompting trends.