Agentic AI Systems, From First Principles
A linear engineering path through the book's agentic design patterns: model boundary, prompt contracts, control flow, tools, memory, protocols, safety, evaluation, prioritization, discovery, and a capstone agent architecture.
How to read this track
Read from 00 to 24. The sequence follows the PDF closely: foundations first, then the 21 design patterns in book order, then a composition lesson that folds in the conclusion and appendices. Each lesson translates the book pattern into engineering state, contracts, failure modes, an implementation sketch, and a running coding/research assistant example.Mental model
An agent is a stateful control loop around a model. The patterns are not separate tricks; they are the named ways to strengthen the loop: context, decision, action, observation, memory, collaboration, recovery, safety, evaluation, and resource control.The dependency ladder
00-02
Foundations Define the agent boundary and prompt/context contracts.
03-08
Execution Build line, branch, parallel work, reflection, tools, and planning.
09-12
State and protocols Add multi-agent roles, memory, learning, and MCP.
13-18
Control and grounding Add goals, recovery, human review, RAG, A2A, and resource budgets.
19-23
Trust and autonomy Add reasoning, guardrails, evals, prioritization, and exploration.
24
Composition Assemble the full coding/research agent and map the framework choices.
Lessons
Part 0 - Map
Part I - Foundations
01
Agent boundary - model, policy, state, environment
Separate the model from the control system around it before adding patterns.
Introduction and Agent characteristics; PDF outline pages 14-23.
02
Prompt and context contracts - the interface to the model
Prompting becomes engineering when inputs and outputs are typed, scoped, and validated.
Appendix A plus Chapter 1 context-engineering discussion; PDF outline pages 250-266 and 24-32.
Part II - Core execution patterns
03
Prompt chaining - linear decomposition
Turn one overloaded request into a sequence of focused transformations.
Chapter 1 - Prompt Chaining; PDF outline pages 24-32.
04
Routing - conditional control flow
When the next step is not always the same, the agent needs a route decision.
Chapter 2 - Routing; PDF outline pages 33-42.
05
Parallelization - fan-out, fan-in, and latency
Run independent work together, then reduce it back into one state.
Chapter 3 - Parallelization; PDF outline pages 43-52.
06
Reflection - critique, repair, and second-pass quality
A feedback loop lets the system inspect and improve output before continuing.
Chapter 4 - Reflection; PDF outline pages 53-62.
Part III - Action and grounding
07
Tool use - function calling as controlled action
Tools let the agent leave language and interact with live systems.
Chapter 5 - Tool Use / Function Calling; PDF outline pages 63-75.
08
Planning - from goal to executable path
Planning lets the agent choose a multi-step path when the path is not known in advance.
Chapter 6 - Planning; PDF outline pages 76-85.
Part IV - State, collaboration, and protocols
09
Multi-agent collaboration - roles, handoffs, synthesis
Multiple agents help when specialization, isolation, or parallel judgment beats one general loop.
Chapter 7 - Multi-Agent Collaboration; PDF outline pages 86-96.
10
Memory management - session, state, long-term knowledge
Memory is scoped storage with write and read policies, not a longer transcript.
Chapter 8 - Memory Management; PDF outline pages 97-110.
11
Learning and adaptation - improving from experience
Agent learning starts with traces, feedback, and system updates before model training.
Chapter 9 - Learning and Adaptation; PDF outline pages 111-119.
12
MCP - model context protocol as integration contract
MCP standardizes how agents discover and use external resources, prompts, and tools.
Chapter 10 - Model Context Protocol; PDF outline pages 120-130.
Part V - Reliability and control
13
Goal setting and monitoring - objective, progress, stop
An agent needs a measurable target and a way to know whether it is getting closer.
Chapter 11 - Goal Setting and Monitoring; PDF outline pages 131-140.
14
Exception handling and recovery - controlled failure
Reliable agents treat errors as expected observations with recovery paths.
Chapter 12 - Exception Handling and Recovery; PDF outline pages 141-146.
15
Human-in-the-loop - approval, correction, escalation
Human judgment belongs at risk, ambiguity, and accountability boundaries.
Chapter 13 - Human-in-the-Loop; PDF outline pages 147-152.
Part VI - Knowledge, communication, and optimization
16
RAG and agentic retrieval - open-book reasoning
Retrieval grounds the agent in external knowledge; agentic retrieval decides when and how to look.
Chapter 14 - Knowledge Retrieval / RAG; PDF outline pages 153-163.
17
A2A - agent-to-agent communication
Remote agents need identity, capability discovery, task lifecycle, and traceable messages.
Chapter 15 - Agent-to-Agent Communication; PDF outline pages 164-174.
18
Resource-aware optimization - cost, latency, budgets
Agents must manage compute, time, money, context, tools, and human attention as scarce resources.
Chapter 16 - Resource-Aware Optimization; PDF outline pages 175-185.
Part VII - Reasoning, safety, and evaluation
19
Reasoning techniques - decomposition, search, verification
Reasoning is controlled extra computation spent to improve hard decisions.
Chapter 17 - Reasoning Techniques plus Appendix F; PDF outline pages 186-198 and 287-298.
20
Guardrails and safety - layered constraints
Safety is built from input checks, tool permissions, output filters, review, and monitoring.
Chapter 18 - Guardrails and Safety Patterns; PDF outline pages 199-209.
21
Evaluation and monitoring - traces, metrics, regressions
Agent evaluation scores the trajectory, not just the final answer.
Chapter 19 - Evaluation and Monitoring; PDF outline pages 210-219.
Part VIII - Autonomy and discovery
22
Prioritization - choosing the next best action
Priority keeps an autonomous agent focused when many useful actions are possible.
Chapter 20 - Prioritization; PDF outline pages 220-227.
23
Exploration and discovery - finding unknown unknowns
The frontier pattern: agents proactively search for information, hypotheses, and possibilities.
Chapter 21 - Exploration and Discovery; PDF outline pages 228-236.