Chapter 10 of my book is about what lasts. Its checklist is about what to do on Monday. The pattern I call constrained delegation is built one layer at a time, not declared all at once. Here is how a team can start this week.
Diagnose before you upgrade
Pick whatever delegated system is misbehaving right now. Before anyone suggests a bigger model, name which of six layers actually failed. The layers are legibility, evals, context, runtime, authority and supervision. The chapter's point is that a legible workplace without evals is not enough, evals without good context are not enough, and so on down the chain.
Hold a thirty-minute review with the people who saw the failure. Go through the six layers in order and ask of each one whether it held. Write the answer in one sentence. In the two recurring cases the book followed, the model was rarely the layer that failed. Your case may differ. The exercise makes you check before you spend.
Set the dial and name the owner
Next, choose one real workflow and set its autonomy dial in writing. Joel Hron's framing, as the chapter uses it, treats agency as a dial. Turn it up where the work is reversible and recoverable. Turn it down where a wrong move costs something you cannot get back. Write down, for that workflow, what the agent may do alone, what needs a human, and why. The chapter puts it sharply: an undocumented dial is a dial nobody actually set.
Then name the person accountable before you ship. Ask who owns this system's output six months from now, after whoever wrote the prompt has moved to a different team. If no name comes to mind, the responsibility was never assigned. Write the name next to the workflow. This takes five minutes and prevents the longest argument you could have later.
Build one thing
Now pick the chapter that maps to your worst gap and build that one thing. The book offers a short menu. It names a legible repo, a real eval slice, a working set instead of a document dump, a durable runtime, a bounded credential, a control plane, or a named owner. Choose one. Scope it to a week. Resist the urge to do three.
Last, put a reminder in your calendar for six months out. Reread the seven "make" lines from the end of the chapter. Make the task legible. Make the standards explicit. Make the context usable. Make the runtime durable. Make authority narrow enough to trust. Make review proportionate to risk. Make the organization capable of learning.
The test is not whether you agreed with the book. The test is whether more of those seven lines are true of your system than are true today. Score yourself now, on paper, so the comparison is honest later.
This post draws on Chapter 10 of From Copilot to Colleague.