How To Build Agentic AI Systems? | Breakdown of Claude Code and ChatGPT

Everyone is building agents. Most of them are workflows wearing a costume. This session works from first principles: an LLM is stateless. Every call is a fresh computation, with no memory of prior interactions unless you explicitly provide it. Everything we call "AI engineering" - memory, retrieval, tools, MCP, the agent loop itself - is scaffolding built around that single constraint. Understand the constraint and the architecture stops being a menu of frameworks and starts being a set of decisions. What this covers: 00:00 Intro — what you'll walk away with 02:31 LLMs are inherently stateless 06:51 Context engineering: crafting what the model sees 12:07 The four levers: memory, retrieval, tools, MCP 23:05 Why you can't just stuff the context window 27:23 The 200K ceiling, and how big that really is 31:52 Managing context: compaction, sliding windows, summarisation 35:08 Coding agents: harness, context, and model all change 37:54 What is an agent? 42:44 Inside Claude Code: the Planner → Generator → Evaluator loop 48:45 Sub-agents and skills at scale 1:05:11 Five core problems in agentic systems 1:12:08 Agents vs workflows: who controls the loop 1:18:11 The real cost — and when an agent is justified *If you'd like to dive deeper into the technical aspects of AI* Checkout: https://123ofai.com *Feel free to connect with us on* LinkedIn:   / 123ofai   Instagram:   / 123ofai   #aidesign #machinelearning #aidevelopment