Agent technology and orchestration

How should these systems be built so results are attributable?

Building an agentic system is not simply a matter of choosing the right model or framework. Architecture determines what the agent can do, how it behaves, and whether its results can be understood and reproduced. This pillar examines model selection and routing, tool use, memory, Model Context Protocol, retrieval, and multi-agent orchestration, with a focus on understanding which component is responsible for which outcome.

We treat orchestration as a design choice with measurable cost and complexity, not as a synonym for intelligence. We examine when a specialist agent adds value, when a simpler workflow is sufficient, and where established methods such as rules, operations research, optimization, and simulation should remain responsible for solving well-defined problems. In many supply chain applications, the language model may be better positioned as an interface, planner, critic, or exception handler than as the numerical solver itself.

The objective is to develop architectures that are modular, observable, reproducible, and attributable, so that agentic AI can be evaluated as an engineered system rather than treated as an opaque black box.