Building Agentic AI — The Big Picture

Preface to a six-part series on warehouse management.

When you use AI as an assistant, you know what you’re doing. You ask, you get an answer. The model produces something, you evaluate the output and make a decision. You are the human in the loop. The AI goes as far as a suggestion, and from there you take over.

Agentic AI changes this loop.

An agent is a system that acts when given a goal. It accesses tools, queries data, makes intermediate decisions, and plans the next step. You don’t have to approve every step; the system moves forward in its own flow. This is a significant difference — because now there’s not an assistant but a decision actor in front of you.

We observe this transition at a global scale as well. Gartner’s latest analyses paint a clear picture: AI agents have moved beyond the experimentation phase. Organizations are no longer asking “Should we try this?” but rather “How do we do this right?” In other words, agents have moved out of the assistant role and into business processes.

But technology alone is not enough to create value. I see this clearly in the field.

An agent built without transforming processes simply becomes a machine that spins existing chaos faster. Every layer added without redesigning ways of working creates new burdens. An agent running without the right context produces wrong decisions. If there is no data quality, no matter how intelligent a model you use, the output will be unreliable. Above all, no architecture stands without organizational readiness.

To make the construction of Agentic AI visible — to show how reliable agents can be built and how they can be brought to life with end-to-end architecture — I planned and wrote a series of articles. I wrote this series not with a theoretical framework, but by starting with a concrete problem: warehouse management.

I know it sounds ordinary. But this is genuinely a perfect testing ground for agents. There is real-time data. There are multi-step decisions. Most importantly, when something goes wrong, the result is visible. Products falling from shelves can be counted; incorrect assignments can be reported. This visibility makes architectural decisions both possible and necessary.

Four agents operate in this system. Here is a table I’d like to share:

Two of the four agents do not use a large language model — and this led me to an important realization.

Agentic AI does not mean a system that uses a large language model.

Three things make an AI system agentic: each agent’s role and responsibilities being clearly defined, agents’ outputs feeding into one another, and the system having consciously designed human integration. Large language models are a tool within this. They may or may not be present.

So where does the human stand in this system?

Answering this question correctly is perhaps the most critical design decision from an architectural standpoint. There are two different models.

• Human-in-the-loop: In this model, the system stops and cannot proceed without human intervention. This model activates at every point where a decision carries a legal commitment, creates a financial obligation, or is irreversible. A concrete example from the warehouse system: overtime approval is a legal commitment. The system cannot — and should not — make this decision on its own.
  

• Human-on-the-loop is different. The system runs and the human monitors. Routine assignments, compliance checks, inventory balance audits — these are handled by the agent, and the human only intervenes in the event of an anomaly. There is no need to approve every step.

  
  

  The choice between these two models is not arbitrary. It is determined by the criticality of the process, the reversibility of the decision, and legal obligations. If it is not done at the design stage, the agent is either over-constrained and its value drops to zero, or it is left too free and loss of control begins.
  

  The six-part series I wrote is a layered unfolding of a whole. Reading in order gives you the big picture; starting from the layer you’re interested in is also possible.
  

• Data Layer: If the data layer is rotten, the agent makes rotten decisions too. The foundation of the entire architecture starts here. (Read more: The Data Layer)
  

• Semantic Layer: Data exists but meaning cannot be produced without generating it first. The foundation of the RAG infrastructure is built in this layer. (Read more: The Semantic Layer)
  

• Ontology & Knowledge Graph: An agent that understands relationships between concepts can think in context. This is the agent’s memory. (Read more: Modeling Relationships with Ontology and Knowledge Graph)
  

• Orchestration: The agent doesn’t make decisions in a single step, but in a chain. The mechanism that makes this chain reliable is orchestration. (Read more: The Brain of Agents)
  

• Observability: A system that cannot be observed cannot be managed. The agent’s nervous system — why is it essential to see what’s happening inside in real time? (Read more: The Agent’s Nervous System)
  

• Audit & Explainability: Trust is not won at the moment of deployment, but re-earned after every decision. How to build an accountable agent? (Article to be published soon)
  

A well-designed agent system does not sideline humans. It empowers them. It enables better questions, faster analysis, and clearer options — but it must never be forgotten that it is always the human who sets strategy, assigns priority, and bears risk.

Agents can be intelligent. Intelligence alone is not enough to be reliable. Reliability is built with accurate data, meaningful context, transparent decisions, and accountable architectures. Layer by layer. Step by step. And always alongside process design.

Dr. Şükrü İmre

Note: The use case design, architectural approach, layer decisions, and human-in-the-loop points in this article series were conceived and defined by me. Claude (Anthropic) was used as a collaborator in code development and article editing. All content has been tested and verified.