Planning-Provisioning Dual-Agent Architecture

Planning-Provisioning Dual-Agent Architecture is a way of turning a person's request into a working system by splitting the job into two stages, each handled by a different AI. Think of a cafe. A customer says, "Something cold and not sweet," and the order-taker turns that into a clear ticket like "Iced Americano, no syrup." The barista then makes the drink using only that ticket. Here, the order-taker who writes up the ticket is the planning role, and the barista who actually makes the drink is the provisioning role.

Splitting the roles this way means people can simply ask in natural language without knowing exact terms or procedures, and they don't have to worry about the complicated process of picking, assembling, and connecting the right components behind the scenes. Because the planning side and the building side are separated, the building side only ever has to process a fixed-format ticket, which cuts down on mistakes and speeds things up.

In an internal tool deployment system called PDI Brew that uses this architecture, the first AI turns an employee's request into a clarified blueprint, and the second AI reads that blueprint and automatically provisions the necessary cloud resources, producing a finished web app in seconds.

The article explains that "PDI Brew operates through a two-stage agent architecture," where the planning agent converts the user's intent into a manifest, and the provisioning agent breaks that down to orchestrate AWS resources. A common misunderstanding is thinking of this as "one smart AI doing everything on its own," but the key point is that it's actually two distinct processes with clearly separated roles: one that clarifies and one that executes.

Next time you ask a chatbot or coding assistant to build an app or a document, don't just say "build it" all at once — split it into two steps. Step 1: "Organize what I want into an itemized plan, but don't execute it yet." Step 2: After reviewing the plan, say "Now execute this plan." Splitting it this way, you'll notice the results become noticeably more accurate.