The Orchestrator

Chapter 1 sketched SaaS onboarding, media redesign, public-service permits, and financial reporting. In stack terms, each one looks much the same: standard connections to the warehouse and the ticket system, a LangGraph or CrewAI fan-out running agents in parallel, a frozen evaluation set, a critic agent that may score results but never publish anything, shared trace IDs across the run, and checkpoints so a run survives a crash — because “the model said the numbers match” is not a control. In every one of those domains, the hard part is deciding who may propose and who must sign — not the interface.

That decision needs a record: a document that accompanies it. When a decision is made by an element affected by an inference engine — by delegated intelligence — the decision record shows how the decision was made and what drove it: what policy applied, which authority stood behind the policy, who is accountable for the decision. If there is a policy that needs to be enforced, the record identifies what the policy is, whose policy it is, and how it was enforced. Earlier in this chapter, the developer and the orchestrator looked at an evidence packet to decide which proposed solution to accept. The decision record is the mirror of that on the decision side: it is what accompanies an outcome or a decision once it has been made.