Multi‑agent LLM development is just a distributed systems problem — AGI won’t magically fix consensus

The pitch and the pushback

A recent blog post argues that building systems of cooperating LLMs should be treated as a distributed systems problem, not a machine‑learning mystique. The author says choreographic languages — concise formalisms for describing interactions — look promising for the bespoke workflows that spring up between agents, especially when you fold in game‑theoretic thinking. It has been reported that some colleagues are skeptical, even defeatist: newer, smarter models will allegedly make languages and verification tooling obsolete. That’s the “AGI will save us” refrain. Cute idea. Probably premature.

Consensus, formalized

Put simply: a natural‑language prompt is underspecified, so there’s a set Φ(P) of programs that satisfy it. When you spawn agents A1…An to implement pieces φ1…φn, they must refine a single coherent φ in Φ(P). That’s just distributed consensus dressed up in Python and API keys. Design choices made by one agent constrain others; library choices, async callbacks, data schemas — they ripple outward. Ask for “an app to track recipes” and you get back a distributed decision problem, not just code generation. Want the short version? Multiple actors, overlapping specs, and the need for agreement — that’s classic distributed systems, full stop.

Why intelligence isn’t a panacea

The post reminds us of a blunt truth from CS theory: some impossibility results are invariant to how smart the participants are. Asynchrony, partial failure, and ambiguous specifications don’t melt away because your agents are clever. You can paper over problems with heuristics and retries, sure — but you can’t conjure consensus where the model of communication makes it impossible. So no, AGI won’t automatically make multi‑agent programming trivial. Tools, languages, and formalism still matter. The author teases a forthcoming paper on a choreographic language for agent workflows; keep an eye out. If you care about reliable multi‑agent systems, expect engineering, not miracles.

Sources: kirancodes.me, Lobsters