Most agent demos work. Most agent systems fail on the same thing: not the model, not the prompt, but the knowledge layer underneath — how the agent stores what it knows, updates it without breaking, retrieves the right thing under load, and composes it into context for the next decision. That layer is where reliability is won or lost in production, and it’s an engineering discipline in its own right.
This is a practitioner series that maps it: RAG pipelines, vector stores, knowledge graphs, agent memory, context construction — what each is actually for, where it breaks, what it costs to run, and how to combine them when one isn’t enough. Named vendors, real numbers, honest tradeoffs. Written for engineers building agent systems, and staged so the people deciding what to build can follow the argument.
I’m Ratko Nikolić, a Senior AI Agents Engineer at Unlimit, where I ship production agent systems; previously Lead AI Engineer at InterVenture, where I built an agentic knowledge-management system over the company’s internal knowledge, a conversational data agent over business data, and GenAI systems for content and product classification. I hold master’s degrees in Computing in the Social Sciences (University of Belgrade) and Computational Finance (RAF).
New articles land here first. Subscribe for the full series. I take on a small number of advisory engagements each year for teams hitting exactly these problems — reach out if that’s you.
