KirkDBorne: agents, memory, multimodal stacks

Enterprise generative AI is moving out of the chatbot phase. That is the real story here. Kirk Borne’s post points people toward Denis Rothman’s book because the hard part is no longer getting a model to answer — it is building a whole system that can remember, route work, use the right model, and stay safe inside a business. The shift is from “pick a frontier model” to “design the stack around the model.” ### What is the actual thing being promoted? It is Denis Rothman’s *Building Business-Ready Generative AI Systems*, published by Packt in July 2025. The book is aimed at people building enterprise AI products, not just experimenting with prompts. Its subtitle is basically the thesis — human-centered systems with context engineering, agents, memory, and LLMs for enterprise use. does “business-ready” matter so much? Because a demo only has to work once. A business system has to work every day, with messy inputs, permissions, latency limits, compliance rules, and users who expect consistency. That is why the book frames a generative AI system as a controller architecture, not a single model call. Packt’s description says the controller needs two key pieces — a conversational agent and an orchestrator. That is a very enterprise way to think about it. ### Why are agents suddenly central? An agent is the part that decides what to do next — answer directly, call a tool, retrieve context, hand work to another component, or ask for clarification. But agents are brittle if they only live inside one prompt window. Rothman’s repo and book materials keep coming back to orchestration, dynamic RAG, and multi-step control, which tells you the goal is repeatable workflows, not clever one-off outputs. ### Why is memory such a big deal? Because enterprise work is mostly context. A useful assistant has to remember prior conversations, company rules, customer history, and task state without stuffing everything into one giant prompt. The book pitches “advanced memory retention” and even “neuroscientific memory systems,” which sounds grandiose, but the practical point is simple — short-term chat memory is not enough for long-running tasks. ### What does multimodal add here? It means the system can reason across more than text — images, documents, maybe structured business data too. In enterprise settings, that matters because real work arrives as PDFs, screenshots, forms, dashboards, and mixed data sources. The book explicitly highlights multimodal reasoning and image generation, which signals a stack built for operational inputs, not just chat transcripts. ### Where do observability and safety fit? They are the part people skip in prototypes and then rediscover in production. If an agent makes a bad tool call or retrieves the wrong document, you need traces, guardrails, moderation, and security layers to see what happened and stop it happening again. The book description explicitly includes secure integration, security, and moderation — which is a clue that the target reader is shipping systems inside real organizations. ### Why mention model selection at all? Because enterprises are no longer betting on one model forever. They mix vendors, costs, latency tiers, and task-specific strengths. One model may be best for extraction, another for reasoning, another for image work. The book’s sales copy even calls out integrating OpenAI and DeepSeek models “as you see fit,” which is really an argument for architecture over allegiance. ### So what changed? The new center of gravity is context engineering. Turns out the scarce resource is not raw model intelligence but the system’s ability to assemble the right context, memory, tools, and controls at the right moment. That is also why Rothman already has a follow-on book about context engineering for multi-agent systems. The market is telling builders that prompt craft was the warm-up act. ### Bottom line? Borne’s post matters less as a book plug than as a signal. The conversation around enterprise AI has matured. The winning question is no longer “which model should we use?” It is “what architecture lets this thing remember, reason, route, and behave reliably at work?”