most multi-agent coding frameworks seem to prefer strict task isolation -- split the work, avoid interference, merge later. humans end up acting as the coordinator.

is this a deliberate architectural choice (isolation is simpler and safer), or are we hitting a real capability ceiling where agents can’t reliably coordinate on a shared, evolving codebase?

the ideal would be multiple agents working on the same system with awareness of each other’s intent and changes -- not just “one agent per file.” is anyone meaningfully solving that problem, or is divide-and-conquer simply good enough for now?