Multi-agent LLM software development is fundamentally a distributed systems problem, not merely an intelligence problem. The author argues that dismissing coordination tooling with 'just wait for smarter models' ignores decades of distributed systems theory. Using formal models, the post maps multi-agent code synthesis to distributed consensus, then applies FLP impossibility (no system can guarantee safety, liveness, and fault tolerance simultaneously) and the Byzantine Generals theorem (consensus requires fewer than 1/3 faulty nodes) to multi-agent workflows. These impossibility results are invariant to model intelligence — smarter agents shrink constants but cannot remove fundamental bounds. Practical takeaways include using failure detectors to improve liveness guarantees and using tests/static analysis to convert byzantine misinterpretations into crash failures. The post concludes that scaling multi-agent development requires deliberately designed coordination protocols, not ad-hoc workarounds.
Sort: