Inside an AI Agent: Orchestrator, Reasoning Core, Memory, Tools | Ep 4 Designing Agentic Systems

Last episode I told you an agent is a program that talks to an LLM. That was a simplification, and today I am taking it back. When you use ChatGPT or Claude you see one package. Instructions in, response out. No agent layer, no reasoning core, no tools. In an enterprise you have to see all of it, because you will be accountable for most of it. An agent is four parts. The orchestrator, which is the program you wrote. The reasoning core, which is the LLM. Memory. And tools. Only one of those four is the AI. The other three are things you have been building for twenty years. In this video: Why the reasoning core can never connect to a tool, a database or anything else Why it has no memory, and what your orchestrator has to re-send on every single call What a tool actually is: an API endpoint, a database, a connector, a cache How the orchestrator maps a plain statement from the model to a real API call The failed 50,000 rupee transfer, walked end to end through all four components Why passing more context costs you real money And the moment the model asks for a reversal, and the orchestrator refuses The key takeaway: the model decides the next step, and it cannot take it. Every call that reaches a bank system is made by your code, with your credentials, after your check. That is why the human in the loop is a design decision and not an afterthought. Next episode: single agent or many. What a team of specialised agents really costs you, and when one agent is still the better answer. I am an Enterprise Architect with 23 years in banking and financial services. This series is about designing agentic AI systems that survive production, audit and a regulator. Not demos. Subscribe for a new episode every week. #AgenticAI #AIAgents #EnterpriseArchitecture #LLM #BankingTechnology #SoftwareArchitecture #AI