Multi-agent patterns that hold up in production (vs the ones that only work in demos)

There are multi-agent patterns that work well in s and fall apart on real work. After running agent teams on production tasks for a few months, here is how I think about it. What actually works Shared state file, not message passing The most reliable coordination pattern: agents do not pass messages back and forth. They read and write a shared state file. Each agent reads current state, does its work, writes results back, and exits. Slower than real-time coordination but far reliable. Agents can be killed and restarted without losing work. You can inspect state at any point.