April 14, 2026

Group project energy, now with AIs

Multi-Agentic Software Development Is a Distributed Systems Problem

Smarter AIs won’t stop the mess — this is still a group project

TLDR: A researcher argues that getting multiple AIs to build software is a coordination problem that needs better tools, not just smarter models. Commenters split between “iterate like humans do” and “you still need an architect,” turning it into a management-versus-magic cage match that matters for how AI will actually build real apps

Today’s hot take: building apps with many AI “agents” isn’t magic, it’s a group project from chaos. The author says it’s basically a consensus problem — like getting a dozen cooks to agree on one recipe — and argues we need new languages to keep the robots in sync. But the comments? Fireworks.

One camp dunks on the “just wait a few months” optimism. Smarter models won’t erase coordination, they say; even humans hit the same walls. Others clap back with the Linux card: if people wrangled a monster codebase, don’t tell us math proves AIs can’t. As falcor84 puts it, no theorem says robots are doomed. The pragmatists bring vibes, not proofs: just ship something, then align. “Build one thing so everyone can see, then iterate,” echoes through the thread. jbergqvist argues a “main agent” can review and fix subagents’ messes faster than going solo.

Meanwhile, the architects storm in with Conway’s Law: org charts shape software. Translation: more bots without a boss equals spaghetti. “Maybe what’s missing is the architect,” says SamLeBarbare. Memes fly: “AGI fairy dust won’t merge your PR,” “LLM group project energy,” and the new proverb: “Two AIs can code anything… twice.” Verdict? The tech is flashy, but the drama is pure management

Key Points

  • The article frames multi-agent LLM software development as a distributed systems coordination problem.
  • The author argues that smarter future models will not eliminate coordination issues due to fundamental impossibility results in distributed systems.
  • A forthcoming choreographic language is mentioned as a concise formalism for describing multi-agent workflows, potentially incorporating game theory.
  • A formal model is presented: natural language prompts are underspecified, and multi-agent synthesis requires consensus on a single consistent interpretation.
  • Task decomposition and parallel agent work necessitate agreement on interdependent design decisions to ensure the composed system satisfies the prompt.

Hottest takes

"there's no math that can help you prove that AIs can't" — falcor84
"what stops the main agent from reviewing the code and sorting out any inconsistencies" — jbergqvist
"Conway’s law still applies" — SamLeBarbare
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.