Volume III

Agentic Machine Learning Systems

Lecture decks for Volume III, which covers machine learning systems whose unit of work is a trajectory, a sequence of model calls, tool actions, and state that unfolds over many steps. Every lecture comes as a PDF for projection, a PowerPoint file whose notes pane holds the speaker notes, and the LaTeX source.

All PDFs (ZIP) All PowerPoint (ZIP) Source

Introduction

Course OverviewEngineering the runtime around a model that actsPDF PowerPoint Source

IntroductionWhy a correct action can still fail the taskRead the chapterPDF PowerPoint Source

Part I: The Model

The Foundation ModelOne model call, its contract, and its costRead the chapterPDF PowerPoint Source

Test-Time ComputeWhen more deliberation is worth its costRead the chapterPDF PowerPoint Source

Part II: Agent Memory

Context EngineeringChoosing what each model call seesRead the chapterPDF PowerPoint Source

KV Cache ManagementHolding attention state across agent turnsRead the chapterPDF PowerPoint Source

Long-Term MemoryStoring and retrieving claims an agent can trustRead the chapterPDF PowerPoint Source

Part III: Tool Use

Tool CallingThe runtime between a tool call and the worldRead the chapterPDF PowerPoint Source

Agent SandboxesLimiting what agent code can reachRead the chapterPDF PowerPoint Source

Agent SecurityTracking where each value came from and goesRead the chapterPDF PowerPoint Source

Part IV: The Agent Runtime

The Agent HarnessThe software that decides whether an agent continuesRead the chapterPDF PowerPoint Source

Durable ExecutionSurviving a crash without repeating an effectRead the chapterPDF PowerPoint Source

Failure RecoveryRepairing a world left half changedRead the chapterPDF PowerPoint Source

Agent EvaluationEvidence that an agent configuration worksRead the chapterPDF PowerPoint Source

Operating AgentsTraces, postmortems, and staged releases in productionRead the chapterPDF PowerPoint Source

Part V: Learning from Trajectories

Trajectory CurationWhich trajectories become training dataRead the chapterPDF PowerPoint Source

Trajectory Fine-TuningTraining a model on agent trajectoriesRead the chapterPDF PowerPoint Source

Reinforcement Learning from Verifiable RewardsLearning from automated checks used as rewardsRead the chapterPDF PowerPoint Source

Part VI: Agents at Scale

Multi-Agent CoordinationWhen splitting a task across agents paysRead the chapterPDF PowerPoint Source

Agent EconomicsThe cost of each accepted taskRead the chapterPDF PowerPoint Source

Part VII: Synthesis

ConclusionClosing invariants across one agent trajectoryRead the chapterPDF PowerPoint Source

Read the matching chapter at mlsysbook.ai/vol3.

Tip

PPTX files are image-based (300 DPI) and visually identical to the PDF. Use them for PowerPoint presenter mode and slide annotations. For editable slides, download the LaTeX source.