Google researchers unlock some truths about getting AI agents to actually work

The researchers found single AI agents do better at sequential tasks, multi-agent workflows at ones that can be done in parallel