they are so reliably inconsistent, but markdown engineering goes a long way towards improving this (eg. we have zero branch name issues like Anthropic claims)

the biggest issue I've seen are these inconsistencies, mistakes, bad assumptions, or side quests compounding with multiple agents and no human-in-the-loop

spent some hours yesterday refining my orchestrator prompt, so much better - tl;dr use the default implementation agent unless the user specifically asks you to use the big/hard multi-agent workflow (which has agent-human loops and break points between, for each phase)