Agent Engineer Course Book

Contents

1. How to use this book
2. Chapter 1. Agent architecture

   Workflows and agents · Tasks that need agents · The augmented LLM and the agent loop · Prompt chaining and gates · Routing · Parallelization: sectioning and voting · Orchestrator-workers · Evaluator-optimizer · Autonomous agents and stopping conditions · The cost of multi-agent systems · Orchestration: should the LLM or code decide? · Choosing the simplest structure first · Code provides the structure; the model provides judgment · Agreement is not verification · Using frameworks

3. Chapter 2. Tools and MCP

   Tools are interfaces for agents · Choosing appropriate tools · Names, descriptions and parameters · Returning meaningful results · Managing tool result size · Errors are recovery instructions · Tools with side effects: preview and commit · MCP components · The 2026-07-28 specification: statelessness and handles · Transports, cancellation and long-running tasks · Tool annotations and trust · Authentication and least privilege · Human approval and security requirements · When tools grow into the hundreds · When models call tools through code · Managing tool changes as contracts

4. Chapter 3. Context, memory and state

   Context is a finite budget · What to include: altitude, tools and examples · Just-in-time retrieval and progressive disclosure · Four layers of information · Compaction and tool result clearing · Structured notes and external memory · Isolating context with subagents · Preventing context pollution · Prompt caching and context layout · Checkpoints and idempotency · Designing resumption after interruption · Handoff for long-running work · Where to keep conversation state · The gap between checkpoints and external side effects · The memory lifecycle · Evaluate memory and compaction policies too

5. Chapter 4. Evaluation and observability

   Basic evaluation terminology · Outcome-based evaluation · Trajectory evaluation · Three types of graders and their combination · Making model graders reliable · pass@k and pass^k · Capability and regression evaluations · Building an evaluation set · Evaluation isolation and reproducibility · Tracing · Analyzing failures stage by stage · Making deployment decisions with evaluation metrics · What to suspect first when scores move · Evaluating conversational agents · Turning operational cases into regressions

6. Chapter 5. Operations, cost and safety

   Put stopping conditions in the system · Layers of budget limits · Detecting loops without progress · Timeouts and deadlines · Retries · Concurrency and overload control · Human approval points · Binding approval to execution content · Blocking mechanisms on the execution path · What guardrails cover · Sandboxes and least privilege · Prompt injection and separation of privileges · Incident response · Partial success and compensating actions · Deploying without breaking active tasks · Expanding autonomy gradually

7. Appendix A. Glossary
8. Appendix B. Self-check answers and explanations
9. Appendix C. Exam day guide
10. Appendix D. Reference documents