Four capabilities, one governing principle
Generative AI synthesizes four abilities analogous to human cognition — creating novel content, reasoning (drawing inferences, finding causal links), understanding context (nuance, history, user preferences), and recommending. Powered by LLMs, these abilities are only as good as the context supplied to them.
Key concept · Why context is king
Without sufficient context an LLM produces two failure modes: hallucinations (plausible but factually wrong) and answers that are generally correct but inapplicable to the specific situation. The book's mortgage case makes it concrete: asked for the maximum debt-to-income ratio, a model answers "43%" from general knowledge — but for an FHA loan in Florida through a specific lender, the true cap depends on the loan program, compensating factors, lender overlays, and state rules. The right answer lives in the intersection of context, not the model's static knowledge.
GenAI's value spans horizontal functions (marketing & sales, customer service, HR, finance, operations/supply chain, IT/development, general productivity) and vertical domains (healthcare, finance, retail, manufacturing). Agentic design patterns — like the Task Delegation Framework (a supervisor/orchestrator decomposing work to specialist worker agents) — are the structured, repeatable solutions for managing context and building implicit guardrails into the architecture.
Seven parts and a loop
An AI agent perceives its environment, makes decisions, and takes actions to achieve goals — with autonomy, reactivity, proactivity, and (optionally) social ability. Systems are either agent-based (a single agent) or multi-agent (multiple specialized agents collaborating under decentralized control). Each agent has an internal structure:
| Component | Role |
|---|---|
| Goals | The desired outcomes, updatable from feedback or changing context. |
| Sense (perception) | Gathers data from the environment — the mechanism for acquiring context (often via MCP). |
| Reason (cognition) | The LLM-powered core that interprets data and makes inferences. |
| Plan | Devises the sequence of steps toward the goal. |
| Act (action) | Executes actions with tools — typically via function calling. |
| Memory | Stores knowledge, past experience, internal state, learned information. |
| Coordinate (optional) | Multi-agent only: aligns with other agents, recommended via the A2A protocol. |
Key concept · The agentic loop & function calling
Agents run a continuous sense → reason → plan → act → sense-the-results feedback loop, updating memory each cycle so they adapt over time. Action is enabled by function calling: developers register tools (name, description, parameter schema); the LLM decides when to call a tool and with what arguments, emitting a structured object (e.g. JSON); the agent's code executes it and feeds the result back into the loop.
This anatomy yields five architectural features: modularity (add/remove agents without a redesign), scalability, adaptability (learning via the loop), multimodal interaction, and collaboration (shared memory / protocols in multi-agent systems).
Seven levels, data to multi-agent
The GenAI Maturity Model is a strategic roadmap: each higher level builds on the ones beneath it, and the level you reach is a direct result of the patterns you choose to master.
Prepare data (data foundation)
Acquire, clean, curate, and govern data — quality, licensing, accessibility. The essential prerequisite.
Select model & prompt/serve
Choose pretrained models, do prompt engineering, serve via APIs; basic tool use through function calling.
Contextual enhancement (RAG)
Fetch external context dynamically to augment prompts — the crucial step toward factual, useful output.
Tuning for specificity
Fine-tune (PEFT/LoRA or full fine-tuning) on domain data to make models agent-ready.
Grounding & evaluation
Link outputs to verifiable sources (citations) and add evaluation for accuracy, bias, and safety.
Single-agent systems
One autonomous agent doing multi-step reasoning, planning, and tool use — requires LLMOps/AgentOps.
Multi-agent systems
Multiple specialized agents that collaborate, coordinate (via A2A), and negotiate on complex problems.
Levels 5 and 6 are the true agentic frontier; the book later expands them into a dedicated Agentic AI Maturity Model. Reaching a target capability lets you reverse-engineer the four or five patterns you actually need — a deliberate, cost-effective engineering path.
Function calling, MCP, and A2A
As systems move beyond standalone prompts, three interoperability layers let models and agents interact reliably.
- Function calling
- The LLM triggers specific registered tools with structured arguments (e.g.
get_credit_score(...)).
- MCP (Anthropic)
- Agent → tools. A standard way to describe, discover, and securely invoke tools/data as interoperable services — vertical integration.
- A2A (Google)
- Agent → agents. A protocol for structured task delegation and collaboration between agents — horizontal integration.
Key concept · How the protocols compose
They work together: an orchestrator uses A2A to delegate to other agents; those agents use MCP to reach the tools and data they need; results flow back through A2A. A2A relies on an agent server (an endpoint exposing the agent's A2A interface, streaming updates via Server-Sent Events) and an agent card (discovery metadata advertising capabilities). The MCP server connects an agent to external tools over a standardized JSON-RPC API. Internally, each agent is an LLM orchestrator + tools/knowledge + memory — its self-contained "brain."
Why promising pilots fail
Moving from experimental PoCs to robust production systems means clearing a web of interconnected challenges across five categories.
| Category | Key considerations |
|---|---|
| Strategic & organizational | Graduating PoCs (ROI, stakeholder alignment), problem–solution fit, change management, user adoption. |
| Data-related | Governance & licensing, data quality, breaking silos, privacy/compliance (GDPR, HIPAA). |
| Model & technical | Robustness & security, drift, secure tool use, scalability, legacy integration, LLMOps/AgentOps, minimizing hallucinations. |
| Resource-related | Acquiring/retaining expertise; compute and development cost constraints. |
| Ethical & responsible AI | Bias mitigation, transparency & explainability, governance frameworks, compliance. |
Case study · The Returns & Order Status agent
A lab-perfect PoC failed within hours in production three ways: poor context management (it kept saying an order was "Shipped," blind to a courier strike), hallucination (it invented a full refund for a Final Sale item outside its knowledge base), and an entangled workflow failure (on failure it just said "contact support" — no ticket, no context handoff). The lesson: a PoC that works on clean data in a lab is not a production-ready system.
Chapter 1 quiz
Fifteen questions on GenAI's capabilities and context, the agent anatomy and loop, the maturity model, the agentic stack, and production challenges. Answer first, then expand Show answer.
Part A · Multiple choice
What is the book's Principle #1?
- A Bigger models are always better
- B Context is king
- C Fine-tune everything
- D Autonomy above all
Show answer
B — Context is king. Reliability depends on equipping the reasoning core with the right context at the right time.
Which is NOT one of GenAI's four core synthesized capabilities?
- A Creating content
- B Reasoning
- C Understanding context
- D Guaranteeing correctness
Show answer
D. The four are creating, reasoning, understanding context, and recommending — correctness is not guaranteed (hence hallucinations).
Which component of an agent enables it to execute actions with tools?
- A Sense
- B Memory
- C Act (via function calling)
- D Goals
Show answer
C — Act. The act component executes planned actions on the environment, typically enabled by function calling.
In the GenAI Maturity Model, which level introduces RAG?
- A Level 0
- B Level 2
- C Level 4
- D Level 6
Show answer
B — Level 2 (Contextual enhancement). L0 is data foundation, L1 is select/prompt/serve, then L2 adds RAG.
Which protocol connects a single agent to its tools and data (vertical integration)?
- A A2A
- B MCP
- C SMTP
- D SSE
Show answer
B — MCP. MCP = agent connects to tools; A2A = agents connect to each other.
What does the A2A "agent card" provide?
- A A billing record
- B Discovery metadata advertising an agent's capabilities
- C The LLM's weights
- D A DLP policy
Show answer
B. The agent card is the discovery layer; the agent server is the endpoint exposing the A2A interface.
Which level of the Maturity Model marks the emergence of true agentic AI with a single autonomous agent?
- A Level 3
- B Level 4
- C Level 5
- D Level 6
Show answer
C — Level 5 (Single-agent systems). Level 6 is multi-agent systems.
In the mortgage case study, why is the model's "43%" answer potentially wrong?
- A The number is mathematically incorrect
- B The specific context (FHA loan, lender overlays, state rules) changes the true cap
- C The model can't do arithmetic
- D DTI ratios don't exist
Show answer
B. 43% is a general QM guideline, but the correct answer depends on the intersection of operational context.
Part B · True or false
A hallucination is an output that sounds plausible but is factually incorrect.
Show answer
True. Insufficient context is a primary driver of hallucinations and contextually-wrong answers.
The "Coordinate" component is required in every agent.
Show answer
False. Coordinate is optional — only for multi-agent systems, recommended via the A2A protocol.
A successful PoC on clean lab data guarantees a production-ready system.
Show answer
False. The Returns agent case shows a lab-perfect PoC failing in production on context, hallucination, and workflow integration.
MCP and A2A are designed to work together in a single workflow.
Show answer
True. An orchestrator delegates via A2A; delegated agents reach tools via MCP; results flow back through A2A.
Part C · Short answer
Name the four stages of the agentic loop.
Show answer
Sense → reason → plan → act (then sense the results as a feedback loop that updates memory).
List the five architectural features enabled by agent anatomy.
Show answer
Modularity, scalability, adaptability, multimodal interaction, and collaboration.
Name three of the five challenge categories for taking GenAI to production.
Show answer
Any three of: strategic & organizational, data-related, model & technical, resource-related, and ethical & responsible AI.