22 IDEAS · CONNECTED ACROSS THE JOURNEY
Follow an idea.
Start with the chapter that introduces a concept, then follow it into implementation and operations.
abstention / abstain
Abstention and missing evidence
Recognize when authorized evidence cannot establish an answer.
Start in Chapter 14 →Explore 16 chapters
- Chapter 14 · 14.1 Run the incomplete and complete paths
- Chapter 1 · Architect's field checklist
- Chapter 2 · 2.4 Qualification questions for the business owner
- Chapter 3 · 3.5 Give the work a stable identity
- Chapter 4 · 4.6 Quality is a release gate, not a rounding adjustment
- Chapter 6 · 6.3 Design the prompt as a contract
- Chapter 7 · 7.3 Judge sufficiency at the level of requested facts
- Chapter 9 · 9.2 Functional capabilities
- Chapter 10 · 10.4 One request across the boundaries
- Chapter 12 · 12.2 Decision record one: retrieval before additional autonomy
- Chapter 15 · 15.4 Separate answer status from review state
- Chapter 16 · 16.3 Use a bounded coordinator
- Chapter 17 · 17.4 Build a handoff packet that another person can use
- Chapter 18 · 18.3 Complete the regression record
- Chapter 19 · Promote an operated system, not a successful demo
- Chapter 22 · 22.10 The ten questions every enterprise AI architect should ask
A2A / Agent2Agent
Agent2Agent protocol
Delegate tasks and exchange status and artifacts between independent agents.
Start in Chapter 16 →Explore 5 chapters
- Chapter 16 · 16.6 Introduce A2A only for a separate task owner
- Chapter 8 · 8.4 A2A delegates to another task owner
- Chapter 22 · 22.4 Decide where RAG, MCP and A2A belong
- Chapter 10 · 10.5 Architecture decisions and alternatives
- Chapter 12 · 12.4 Decision record three: an API boundary and an MCP boundary
governance / risk acceptance
AI governance
Connect permitted use, accountable owners, controls, and release decisions.
Start in Chapter 11 →Explore 4 chapters
AWS / Fargate
AWS production and migration
Connect application promotion to identity, state, recovery, and operational ownership.
Start in Chapter 19 →Explore 9 chapters
- Chapter 19 · Promote an operated system, not a successful demo
- Chapter 1 · 1.3 The case study as the book's spine
- Chapter 4 · 4.7 Stage the investment
- Chapter 9 · 9.9 Define acceptance gates
- Chapter 10 · 10.6 The deployment journey and environment boundary
- Chapter 17 · The requirement that shaped the experimental environment
- Chapter 18 · A release decision needs more than a successful answer
- Chapter 21 · Reconcile the promise with the work people actually complete
- Chapter 22 · 22.11 A closing reflection from the Question Assistant journey
chunking / chunk
Chunking and ingestion
Turn source documents into retrievable units while preserving scope and qualifications.
Start in Chapter 13 →Explore 11 chapters
- Chapter 13 · 13.3 Chunk around the meaning being tested
- Chapter 5 · 5.6 Make the readiness decision explicit
- Chapter 7 · 7.1 Understand the two lifecycles
- Chapter 9 · 9.2 Functional capabilities
- Chapter 10 · 10.1 Architecture requirements and principles
- Chapter 11 · 11.6 Turn the threat model into acceptance evidence
- Chapter 14 · 14.7 Compare improvements at the failing layer
- Chapter 15 · 15.4 Separate answer status from review state
- Chapter 16 · Three roles addressed three different questions
- Chapter 19 · 19.1 Choose a complete reference topology
- Chapter 20 · 20.5 Build a cost ledger with a consistent unit
context engineering / context budget
Context engineering
Select the evidence, instructions, and workflow state a model needs.
Start in Chapter 6 →Explore 1 chapters
data readiness / source inventory
Data readiness
Decide which sources can become current, authoritative, and accessible evidence.
Start in Chapter 5 →Explore 11 chapters
- Chapter 5 · 5.1 Complete a source inventory
- Chapter 1 · 1.3 The case study as the book's spine
- Chapter 2 · 2.3 A practical prioritization model
- Chapter 3 · 3.3 Follow the manual workflow
- Chapter 9 · 9.1 Start with actors and outcomes
- Chapter 10 · 10.10 Side example: a product RFP assistant
- Chapter 11 · 11.1 Assign each control a decision and an owner
- Chapter 13 · 13.6 Update and revoke without leaving hidden copies
- Chapter 14 · 14.5 Treat Q-018 as a useful result
- Chapter 20 · A correct answer can become unsafe to reuse
- Chapter 22 · 22.3 A practical portfolio progression
evaluation / recall
Evaluation and recall
Measure retrieval coverage, factual support, completeness, and useful abstention separately.
Start in Chapter 14 →Explore 18 chapters
- Chapter 14 · 14.6 Build a real evaluation set beyond the demonstration
- Chapter 6 · 6.7 Evaluation: behavior, not vibes
- Chapter 1 · 1.5 The architect as translator—and challenger
- Chapter 2 · 2.1 Map the workflow before scoring the use case
- Chapter 4 · 4.3 A transparent unit-economics model
- Chapter 5 · 5.6 Make the readiness decision explicit
- Chapter 7 · 7.5 Define metrics before comparing designs
- Chapter 9 · 9.7 Trace requirements into evidence
- Chapter 10 · 10.5 Architecture decisions and alternatives
- Chapter 12 · 12.2 Decision record one: retrieval before additional autonomy
- Chapter 13 · 13.3 Chunk around the meaning being tested
- Chapter 16 · 16.2 Follow a deliberately bad candidate
- Chapter 17 · 17.2 Make the local run reproducible first
- Chapter 18 · A release decision needs more than a successful answer
- Chapter 19 · 19.3 Build once and identify every behavior dependency
- Chapter 20 · 20.2 Runbook for a revoked source
- Chapter 21 · 21.5 Avoid double counting human effort
- Chapter 22 · Opening — one project is evidence, not an enterprise strategy
provenance / lineage
Evidence and provenance
Trace a generated claim back to a specific source version and passage.
Start in Chapter 13 →Explore 7 chapters
- Chapter 13 · 13.2 Separate extraction from evidence publication
- Chapter 2 · 2.6 Adjacent examples: where this pattern transfers—and breaks
- Chapter 5 · Decide what may become evidence
- Chapter 7 · 7.4 Choose more elaborate patterns for demonstrated needs
- Chapter 11 · 11.1 Assign each control a decision and an owner
- Chapter 14 · 14.5 Treat Q-018 as a useful result
- Chapter 22 · 22.2 Distinguish product reuse from pattern reuse
Extractor / Validator
Extractor, Validator, Evaluator
Separate proposing an answer, checking claims, and assessing completeness.
Start in Chapter 16 →Explore 9 chapters
- Chapter 16 · Three roles addressed three different questions
- Chapter 3 · The case and the teaching example
- Chapter 6 · 6.8 Where the case study fits
- Chapter 7 · Retrieve the evidence needed for the question
- Chapter 10 · 10.4 One request across the boundaries
- Chapter 11 · 11.1 Assign each control a decision and an owner
- Chapter 12 · A decision is useful when someone can revisit it
- Chapter 14 · 14.4 Separate structural checks from factual judgments
- Chapter 20 · 20.3 Handle provider outages without inventing evidence
human review / human-in-the-loop
Human review and approval
Bind accountable decisions to the exact answer version and permitted action.
Start in Chapter 15 →Explore 20 chapters
- Chapter 15 · 15.5 Bind approval to the answer being released
- Chapter 1 · 1.5 The architect as translator—and challenger
- Chapter 2 · 2.2 Segment the candidate opportunities
- Chapter 3 · 3.1 Two questions expose the whole problem
- Chapter 4 · 4.5 Sensitivity matters more than a single forecast
- Chapter 5 · 5.2 Establish the minimum information contract
- Chapter 6 · 6.6 Defend the instruction boundary
- Chapter 8 · Name the responsibility before choosing the mechanism
- Chapter 9 · 9.2 Functional capabilities
- Chapter 10 · 10.1 Architecture requirements and principles
- Chapter 11 · The authorization decision behind one answer
- Chapter 12 · 12.1 Separate constraints from preferences
- Chapter 13 · 13.2 Separate extraction from evidence publication
- Chapter 14 · A relevant passage can still leave half the question unanswered
- Chapter 16 · 16.1 Give every role the information needed to disagree
- Chapter 18 · A release decision needs more than a successful answer
- Chapter 19 · 19.1 Choose a complete reference topology
- Chapter 20 · 20.3 Handle provider outages without inventing evidence
- Chapter 21 · 21.1 Keep the measurement contract unchanged
- Chapter 22 · 22.1 Extract the reusable pattern
hybrid retrieval / hybrid search
Hybrid retrieval
Combine lexical matches and semantic candidates for enterprise evidence retrieval.
Start in Chapter 7 →Explore 4 chapters
idempotency / idempotent
Idempotency and durable work
Resume interrupted work and prevent repeated effects when requests retry.
Start in Chapter 15 →Explore 8 chapters
- Chapter 15 · 15.1 Give each interface one responsibility
- Chapter 8 · Name the responsibility before choosing the mechanism
- Chapter 9 · 9.3 Nonfunctional requirements are product requirements
- Chapter 10 · 10.1 Architecture requirements and principles
- Chapter 12 · 12.6 Keep logical responsibilities separate from products
- Chapter 17 · The requirement that shaped the experimental environment
- Chapter 19 · 19.2 Make the migration inventory actionable
- Chapter 20 · 20.3 Handle provider outages without inventing evidence
LLMOps / CI/CD
LLMOps and release management
Version and validate the behavior dependencies of an AI application.
Start in Chapter 19 →Explore 9 chapters
- Chapter 19 · 19.3 Build once and identify every behavior dependency
- Chapter 6 · 6.8 Where the case study fits
- Chapter 10 · 10.6 The deployment journey and environment boundary
- Chapter 11 · 11.7 The governance record that survives deployment
- Chapter 12 · 12.5 Decision record four: prompts as managed configuration
- Chapter 14 · 14.6 Build a real evaluation set beyond the demonstration
- Chapter 17 · 17.4 Build a handoff packet that another person can use
- Chapter 18 · 18.7 Make the stop decision explicit
- Chapter 20 · 20.4 Diagnose silent quality drift
MCP / Model Context Protocol
Model Context Protocol
Expose tools and context through a reusable client-server integration boundary.
Start in Chapter 15 →Explore 8 chapters
- Chapter 15 · One capability, several entrances
- Chapter 8 · 8.2 MCP connects capabilities to compatible hosts
- Chapter 12 · 12.4 Decision record three: an API boundary and an MCP boundary
- Chapter 22 · 22.4 Decide where RAG, MCP and A2A belong
- Chapter 3 · The case and the teaching example
- Chapter 10 · 10.4 One request across the boundaries
- Chapter 16 · 16.6 Introduce A2A only for a separate task owner
- Chapter 19 · 19.1 Choose a complete reference topology
observability / telemetry
Observability and operations
Detect, diagnose, and recover from infrastructure and silent quality failures.
Start in Chapter 20 →Explore 11 chapters
- Chapter 20 · 20.2 Runbook for a revoked source
- Chapter 10 · 10.8 Design for observability
- Chapter 19 · 19.7 Rehearse rollback across application and data
- Chapter 8 · 8.1 Distinguish a workflow from an agent
- Chapter 9 · 9.3 Nonfunctional requirements are product requirements
- Chapter 11 · 11.1 Assign each control a decision and an owner
- Chapter 12 · 12.6 Keep logical responsibilities separate from products
- Chapter 15 · 15.7 Distinguish prompt administration from a new administrative Skill
- Chapter 17 · 17.3 Define what the labs environment contains
- Chapter 18 · 18.1 Give each gate a different question
- Chapter 22 · 22.1 Extract the reusable pattern
prompt injection / prompt-injection
Prompt injection
Contain malicious instructions embedded in documents and tool responses.
Start in Chapter 11 →Explore 9 chapters
- Chapter 11 · 11.5 Control actions even when content screening fails
- Chapter 2 · 2.6 Adjacent examples: where this pattern transfers—and breaks
- Chapter 3 · 3.2 The small source corpus
- Chapter 5 · 5.1 Complete a source inventory
- Chapter 6 · 6.6 Defend the instruction boundary
- Chapter 10 · 10.7 Threat and failure boundaries
- Chapter 13 · 13.4 Inspect the transformation
- Chapter 16 · 16.2 Follow a deliberately bad candidate
- Chapter 22 · 22.1 Extract the reusable pattern
RAG / retrieval-augmented generation
Retrieval-augmented generation
Select relevant authorized evidence and supply it to a model at answer time.
Start in Chapter 7 →Explore 10 chapters
- Chapter 7 · Retrieve the evidence needed for the question
- Chapter 22 · 22.4 Decide where RAG, MCP and A2A belong
- Chapter 1 · Opening: Who owns the last mile?
- Chapter 2 · 2.2 Segment the candidate opportunities
- Chapter 3 · The case and the teaching example
- Chapter 9 · Opening: “Answer accurately” is not a requirement
- Chapter 11 · The authorization decision behind one answer
- Chapter 12 · A decision is useful when someone can revisit it
- Chapter 14 · A relevant passage can still leave half the question unanswered
- Chapter 16 · Three roles addressed three different questions
ROI / return on investment
ROI and benefits realization
Compare the investment hypothesis with adoption, effort, cost, and realized benefit.
Start in Chapter 21 →Explore 3 chapters
Skill / Skills
Skills and plugins
Package workflow guidance and integrations without replacing backend authority.
Start in Chapter 8 →Explore 5 chapters
- Chapter 8 · 8.3 Skills guide work and packages distribute it
- Chapter 15 · 15.7 Distinguish prompt administration from a new administrative Skill
- Chapter 3 · The case and the teaching example
- Chapter 10 · 10.4 One request across the boundaries
- Chapter 12 · A decision is useful when someone can revisit it
UAT / user acceptance testing
User acceptance testing
Verify how intended users complete real workflows before a release decision.
Start in Chapter 18 →Explore 13 chapters
- Chapter 18 · 18.2 Distinguish demonstration fixtures from acceptance data
- Chapter 9 · 9.4 Acceptance scenarios as executable thinking
- Chapter 11 · 11.6 Turn the threat model into acceptance evidence
- Chapter 1 · 1.1 The mandate: business outcome to operating capability
- Chapter 4 · 4.5 Sensitivity matters more than a single forecast
- Chapter 5 · 5.3 Distinguish policy, exception, and operational fact
- Chapter 10 · 10.6 The deployment journey and environment boundary
- Chapter 12 · A decision is useful when someone can revisit it
- Chapter 16 · 16.6 Introduce A2A only for a separate task owner
- Chapter 17 · The requirement that shaped the experimental environment
- Chapter 19 · Promote an operated system, not a successful demo
- Chapter 21 · Reconcile the promise with the work people actually complete
- Chapter 22 · 22.2 Distinguish product reuse from pattern reuse