From Anatomy of an AI AgentOpen the slide →The six parts of every agent, pulled apart so you can see what each one does.
Plates · 19 figures
Explanations you can run.
Each plate takes one mechanism from the text and lets you operate it: step the loop, fill the window, break the cache, route the token. They live inside the chapters; this page collects them.
Architecture and System Design
From The Agent Loop PatternOpen the slide →Call the model, run the tool it asks for, append the result, repeat until done or capped.
From Context Engineering and MemoryOpen the slide →Every token in the window is spent on something; watch what fills it and what compaction buys back.
From Reliability: Retries, Idempotency, and Failure HandlingOpen the slide →Retries that wait the same time arrive together; randomised waits spread the load out.
Context Engineering
From How Context Goes Wrong: Rot, Pollution, and ConfusionOpen the slide →As the window grows, the model's grip on any one fact loosens, worst in the middle.
From The Four Moves: Reduce, Offload, Isolate, RetrieveOpen the slide →Four moves that keep a long-running agent's window small and useful.
Agentic Memory
From How Memory Is Built and RetrievedOpen the slide →Each run deposits a layer; retrieval scores recency, importance and relevance to cut through them.
Agentic RAG
From From Naive RAG to Agentic RAG: The EvolutionOpen the slide →One-shot retrieval misses; an agent grades what came back, rewrites the query, and retrieves again.
Multi-Agent Systems
From Single Agent versus Multi-Agent: The Core TradeoffOpen the slide →Workers that only gather merge cleanly; workers that each decide something produce parts that do not fit.
From The Case For Multi-Agent: Anthropic's Research SystemOpen the slide →A lead agent fans a task out to workers with their own context windows and gets back only summaries.
Building Coding Agents and Harnesses
From Tokens, Context Windows, and the Quadratic ProblemOpen the slide →Every turn resends the whole history, so total tokens processed grow with the square of the turns.
From Prompt Caching and Context ManagementOpen the slide →Append and the cache keeps hitting; edit anything early and everything after it is paid for again.
From Sandboxing, Approvals, and CheckpointsOpen the slide →Every command the agent proposes is allowed, held for approval, or denied before it runs.
Foundational Modelling
From The Transformer, the Engine Inside Every Modern ModelOpen the slide →Each token scores every other token and takes a weighted blend of what they carry.
From LoRA, Fine-Tuning Giant Models on a BudgetOpen the slide →Freeze the big matrix and learn the change as the product of two thin ones.
From Mixtral and Mixture of Experts, More Brain, Same SpeedOpen the slide →A router sends each token to two of eight experts, so most of the model sits idle for any one token.
Planning and Reasoning
From ReAct, Reasoning and Acting TogetherOpen the slide →Reasoning alone hallucinates, acting alone flails; interleaving them grounds each step.
From Let's Verify Step by Step, Rewarding Good ReasoningOpen the slide →An outcome grader only sees the end; a process grader catches the step that went wrong.
Benchmarks
From Chatbot Arena, Letting Real People Pick the WinnerOpen the slide →People pick the better of two anonymous answers; the votes settle into a ranking.