We've been experimenting with different ways to structure agent workflows.

The key is keeping agents focused. Each agent has a single job. When they need to collaborate, they pass context, not control.

Patterns that work

Sequential chains for linear tasks. Parallel agents for independent work. Hierarchical when you need oversight.

Most teams start with sequential. That's fine. Add complexity only when you need it.

The real win is when agents remember what worked and reuse those patterns automatically.