What is happening?
Several current projects are treating agent work as a sequence of distinct responsibilities. A model or system proposes a plan, an execution harness performs tool calls, and a verification layer checks whether the result satisfies a known condition.
Why people care
The split makes failure easier to name. A bad plan, a broken tool call and an incorrect result are different problems, and they need different ways to inspect and recover.
What to watch
The important question is whether the boundaries are real in code: can a user review the plan, see the actions, and inspect the evidence for the final decision? Architecture diagrams alone are not enough.