The point is not that multi-agentic software development is not possible — people are certainly shipping real software with agents today, and they will probably be shipping more slop tomorrow. The argument I am making is more that these people are currently doing so by implicitly resolving these coordination problems, usually through ad-hoc mechanisms with no thought out guarantees or failure modes. The distributed systems literature has precise formalisms for all of these mechanisms, forty years of theorems about what they can and cannot solve for you, and clear rules for when and where different techniques should be applied.

The key takeaway I got from this is that someone still needs to act as the coordinator/orchestrator for these. Maybe that coordination is the future of software development at the every day engineering level, monitoring and correcting agents when they go off trail.