How AI systems work,from input to production.How AI products work,one mechanism at a time.
Thirteen visual chapters about system design, context, retrieval, tools, agents, evaluation, and safety. Connect the mental model to the implementation.Thirteen visual chapters about what AI products can use, what they can do, and how people design, test, and control them.
Understand the systemaround the model.See what happensbehind the answer.
Each chapter begins with the durable mechanism, then shows a concrete system, its failure modes, and the implementation choices that matter.Each chapter starts with a simple explanation, then uses an example to show what changes, what stays the same, and where the idea can fail.
Choose the mechanism you need.Choose the question you want answered.
FoundationsStart with the task
1Inputs and retrievalWhat the AI can use
4Cache tokensCache tokens
How identical prompt beginnings skip repeated work.How an AI service can reuse the work of reading repeated instructions while still writing a new answer.
Context windowsContext windows
How a finite context window limits what a model can use.Why an AI can use only the input available for the current answer—and how older details can be omitted.
EmbeddingsEmbeddings
How learned vectors make semantic similarity measurable.How words, products, and pictures become numerical representations that make related items easier to compare.
RAGRAG
How retrieval adds relevant evidence before generation.How an AI product looks up useful source material before it answers instead of relying only on what the model already knows.
Tools and agentsHow AI systems take action
5Tool callsTool calls
How models request functions that application code controls.How an AI can request a calendar lookup, database search, or other action while the surrounding app validates and executes it.
Agent loopsAgent loops
How an application repeats model decisions, actions, and checks.How an AI agent handles a longer job by taking one step, checking what happened, and deciding what to do next.
Coding agentsCoding agents
How an agent inspects a repository, edits it, and verifies the result.How a coding agent reads a software project, makes a bounded change, and gathers evidence that the change belongs there.
HarnessesHarnesses
The runtime that supplies context, tools, policy, state, and stopping rules.How the software around a model supplies tools, saved state, permissions, and stopping rules for an agent.
MCPMCP
How MCP standardizes connections between AI applications and external services.How a shared communication standard helps AI apps connect to outside tools without making every integration from scratch.
Quality and safetyHow teams test and control them
3EvalsEvals
How repeatable tests compare versions of an AI system.How teams give different versions the same realistic test so an impressive demo is not mistaken for real improvement.
GuardrailsGuardrails
How layered controls reduce specific risks and limit consequences.How several layers of rules, permissions, checks, and human decisions can reduce harm without promising perfection.
TracingTracing
How traces connect the operations behind one request.How a timeline of the steps behind one answer helps people find delays, failures, and surprising decisions.
Products change. The useful questions are stable.Products change. The useful questions stay the same.
Ask what entered context, which component chose the action, what executed it, which policy applied, and what evidence shows the result was acceptable.Ask what information the product received, who allowed the action, what did the work, which limits applied, and how anyone checked the result.
Study the runtime around the modelSee the system around the AI