When William Waites posted on John Carlos Baez's Azimuth blog, he wasn't pitching another orchestration framework. He was introducing 'plumbing' — a statically typed language for composing LLM agents, grounded in symmetric monoidal category theory — and making the case that the whole field has been building on sand.
The argument lands differently than most academic critiques of production AI tooling. Waites isn't just calling LangGraph, CrewAI, and n8n ugly. He's identifying a specific failure mode: without formal structural guarantees, multi-agent systems fail in ways you only discover after spending compute. To make that economic cost concrete, he cites a Forbes investigation finding that heavy Claude Code users can burn through compute equivalent to $5,000/month at retail API rates — costs that companies like Cursor pay close to in reality. Bad orchestration doesn't just waste engineering time; it burns real money.
Plumbing's answer is to move verification upstream. The language is built around a small set of typed primitives: channels carrying streams of typed values, and four structural morphisms — Copy, Discard, Merge, and Barrier — composable sequentially or in parallel. Every composition must satisfy type constraints, which means the compiler can catch mismatched agent interfaces before any LLM calls are made, checking for well-formedness, potential deadlocks, and structural soundness at compile time rather than at runtime.
The practical examples are concrete. Waites walks through an adversarial document composition pattern — a writer agent coupled with a critic — and a multi-agent debate ensemble where temperature can be modulated at runtime through dedicated control ports. Both are expressed in concise, point-free syntax. More importantly, there's a working compiler and runtime behind them; this isn't a whitepaper.
Waites is not working alone. The project draws on collaborations with mathematicians Tobias Fritz and Vincent Danos, which explains why the category-theoretic foundations feel rigorous rather than cosmetic. Symmetric monoidal categories have a long history in programming language theory and process algebra — plumbing is applying decades of existing formal methods to a domain that has largely ignored them.
Whether that matters to practitioners currently reaching for LangGraph is an open question. The orchestration landscape is crowded, and most teams aren't hiring category theorists. But the cost pressure is real, reliability complaints are loud, and Waites has promised a follow-up post tackling more complex coordination scenarios. If the compiler catches your agent topology bugs before they hit the API bill, that's a concrete pitch even without the math.