Series · 14 of 14 parts · Advanced
AI System Design
How production AI agents actually work: the reasoning loop, memory, tools, guardrails, evaluation, and the unglamorous engineering that keeps them reliable.
What you’ll understand
By the end you can design, evaluate, and debug a production-grade AI agent system from scratch.
Episodes
01The ReAct Loop: How Every AI Agent Actually WorksThought, action, observation: the loop that turned one-shot text generators into agents that can check their work.10 min→02Multi-Agent Orchestration: Managers, Workers, and PipelinesWhy one agent isn't enough, and the coordination patterns that keep a team of agents from deadlocking.13 min→03Agent Memory: The Four Kinds and How They Fit TogetherLLMs forget everything between calls. In-context, external, episodic, and semantic memory fix that.8 min→04Tool Calling and MCP: How Agents Touch the WorldTools give a text model hands. MCP makes every tool speak the same language.13 min→05Prompt Engineering for Agents, Production GradeTen techniques that turn a vague instruction into an agent that behaves the same way every run.13 min→06RAG vs Knowledge Graphs: Two Kinds of AI Memory LookupSemantic search answers fuzzy questions. Graph traversal answers precise ones. Real systems need both.9 min→07Hallucination Detection: Trust, but VerifyConfident, plausible, wrong. The detection layers that stop invented facts from reaching users.10 min→08Context Window Management for Big JobsChunking, map-reduce, and retrieval: how agents handle work that never fits in one context window.14 min→09Self-Improving Loops: Agents That Learn From OutcomesTrack outcomes, mine the wins, update the strategy, guard the update. RL's shape without the gradients.11 min→10Guardrails: Input, Process, OutputPrompt injection, tool authorization, leak prevention: the three-layer safety system every agent needs.11 min→11Evaluating Agents: Golden Sets, Judges, and Shadow ModeSame input, different outputs. How to measure whether a nondeterministic system actually works.11 min→12Debugging Agent Systems: Traces, Replays, Root CausesAgent failures raise no errors. The logging, tracing, and replay toolkit that makes them debuggable.11 min→13Entity Resolution: One Thing, Many NamesSam K., Samuel Kim, sam@acme.com. Deciding what's the same thing is harder and more important than it looks.11 min→14Graph Databases for AI: When Relationships Are the DataWhy multi-hop questions crush SQL joins, and how a graph becomes an agent's long-term memory.8 min→