A recent Hacker News post argues that coordinating large language models (LLMs) to write software is fundamentally a distributed systems issue, not a problem that future AGI will simply erase. The author sketches a choreographic language for multi‑agentic workflows and notes that, once agents split a prompt into tasks, they must agree on a single coherent implementation.

Formally, the post defines Φ(P) as the set of programs that satisfy an underspecified natural‑language request P, then describes the consensus condition C(φ1…φn) requiring every agent’s fragment to refine a common φ∈Φ(P). This mirrors classic distributed consensus where design choices—such as a networking library—propagate constraints to other modules, forcing agents to coordinate like nodes in a Paxos‑style protocol.

The author warns against the complacent refrain that “waiting for smarter models” will solve coordination, pointing to impossibility results such as the FLP theorem that hold regardless of agent intelligence. By treating multi‑agent software synthesis as a distributed consensus problem, researchers can leverage decades of fault‑tolerance work instead of chasing fleeting model upgrades. The post concludes that robust tooling, not speculative AGI, will determine scalability.