Dibia separates two families. Workflow patterns, where control is explicit and written in advance: sequential, conditional, parallel. And autonomous patterns, where control emerges: one agent plans, another takes a handoff, several talk until they converge on something. Half the book is code and the other half is judgement about which to pick.
What I took from it
Always start with the explicit workflow. Autonomy earns its place when the decomposition of the task cannot be known before you begin. If it can be known, writing it down is cheaper, more measurable and far easier to repair when it breaks. That sounds obvious, and it is the opposite of what is being built almost everywhere right now.
Termination patterns. Deciding when the system stops is part of the design, not an implementation detail: without that decision, two polite agents will thank each other until the budget runs out.
And the handoff to a person treated as a designed pattern rather than an emergency plan. The question it left me, which I now ask before any automation: when this does not know what to do, who receives it, with what information, and within what time.
What I left out
The code. This is self-published and tightly bound to the libraries of the moment, so that part will age before anyone finishes reading it. There is also a predictable bias: whoever built an agent framework tends to see problems that agents solve. The chapter that asks whether you need a framework at all is, fittingly, the most useful one.