Home / Figures
The book’s diagrams
Every figure here is a diagram drawn for AI Agents, Engineered: one visual language, from the agent loop to the regression gate. All 96 are free to reuse under CC BY 4.0; each page gives you the SVG and a ready-made attribution line, “Figure N.M from AI Agents, Engineered by Enrique Gutiérrez, CC BY 4.0.”
Part I Foundations
8 diagrams
-
Figure 1.1 · Chapter 1 · free
The book’s compass
-
Figure 1.2 · Chapter 1 · free
Chatbot, workflow, and agent differ in one thing only: who decides the next step
-
Figure 1.3 · Chapter 1 · free
Autonomy is a dial, not a switch
-
Figure 1.4 · Chapter 1 · free
The escalation ladder
-
Figure 2.1 · Chapter 2 · free
Autoregressive generation, the engine’s whole method
-
Figure 2.4 · Chapter 2 · free
The anatomy of a chat prompt
-
Figure 2.5 · Chapter 2 · free
Function calling, step by step
-
Figure 2.6 · Chapter 2 · free
Compounding error, drawn
Part II Building Agents
15 diagrams
-
Figure 3.2 · Chapter 3
One pass through the loop has four beats circling a growing history
-
Figure 3.3 · Chapter 3
Three schedules for a model with tools
-
Figure 3.4 · Chapter 3
A minimal agent has exactly four parts, and all four live in the harness—the frame here
-
Figure 3.5 · Chapter 3
The loop has three ways out, and they are not equally trustworthy
-
Figure 4.2 · Chapter 4
Two ways to plan
-
Figure 4.3 · Chapter 4
Self-consistency
-
Figure 4.4 · Chapter 4
Generate, check, revise—a loop only as good as the signal it feeds back
-
Figure 5.2 · Chapter 5
The same scheduling job in two action spaces
-
Figure 5.3 · Chapter 5
The anatomy of a tool across the boundary between your process and the model’s context
-
Figure 5.4 · Chapter 5
Code execution as a universal action
-
Figure 5.5 · Chapter 5
The three ways to reach another system, drawn as a ladder
-
Figure 6.2 · Chapter 6
Progressive disclosure
-
Figure 6.3 · Chapter 6
The three layers this chapter keeps apart
-
Figure 6.4 · Chapter 6
The multiplication a protocol removes
-
Figure 6.5 · Chapter 6
The five internal shapes a skill can take, each paired with the household analogy that names it
Part III Context Engineering
9 diagrams
-
Figure 7.1 · Chapter 7
The anatomy of an assembled context
-
Figure 7.3 · Chapter 7
A bedside chart for context failures
-
Figure 7.4 · Chapter 7
The four operations, arranged around the desk
-
Figure 8.2 · Chapter 8
Retrieval-augmented generation
-
Figure 8.3 · Chapter 8
The quality funnel of a mature retrieval pipeline
-
Figure 8.4 · Chapter 8
One-shot retrieval versus an agentic loop
-
Figure 8.5 · Chapter 8
Three ways to give a model knowledge
-
Figure 9.2 · Chapter 9
Agent memory arranged by temperature
-
Figure 9.3 · Chapter 9
The lifecycle of one remembered thing
Part IV Patterns and Choices
11 diagrams
-
Figure 10.1 · Chapter 10
The augmented LLM, the base unit every pattern in this chapter arranges
-
Figure 10.2 · Chapter 10
Prompt chaining
-
Figure 10.3 · Chapter 10
Routing
-
Figure 10.4 · Chapter 10
The two flavors of parallelization
-
Figure 10.5 · Chapter 10
The evaluator–optimizer loop, the one shape whose defining arrow points backward (in accent)
-
Figure 11.1 · Chapter 11
The orchestrator–worker pattern
-
Figure 12.2 · Chapter 12
The four consequence tiers, keyed to what an action costs when it is wrong and never to the model’s confidence
-
Figure 12.4 · Chapter 12
The autonomy dial
-
Figure 13.3 · Chapter 13
The outer loop: five verbs around durable state
-
Figure 13.4 · Chapter 13
Many loops, one shared store
-
Figure 14.1 · Chapter 14
The ladder of Chapter 1, redrawn as a staircase and priced
Part V Making Agents Reliable
14 diagrams
-
Figure 15.2 · Chapter 15
One agent run drawn as a span waterfall
-
Figure 15.3 · Chapter 15
Quality drift across many runs
-
Figure 15.4 · Chapter 15
Record and replay across the deterministic–nondeterministic seam (the dashed line): your code on the left, the model and the world on the right
-
Figure 15.5 · Chapter 15
The trace-to-dataset flywheel, the loop that turns firefighting into improvement
-
Figure 16.1 · Chapter 16
One agent, two honest summaries of the same ten attempts, drawn here for an illustrative per-attempt success rate of 90 percent
-
Figure 16.2 · Chapter 16
The cheapest eval you can run this afternoon
-
Figure 16.4 · Chapter 16
Eval-driven development as plumbing
-
Figure 17.2 · Chapter 17
The lethal trifecta
-
Figure 17.3 · Chapter 17
Guardrails sort into four families by where they stand
-
Figure 17.5 · Chapter 17
Why a tool’s description is attack surface
-
Figure 18.1 · Chapter 18
One tool failure, two fates
-
Figure 18.2 · Chapter 18
The circuit breaker as a state machine, not a mood
-
Figure 18.3 · Chapter 18
Receipts, not transcripts, are the system of record
-
Figure 18.4 · Chapter 18
The harness grid
Part VI Shipping and Operating
8 diagrams
-
Figure 19.2 · Chapter 19
Why an agent’s cost compounds rather than adds
-
Figure 19.3 · Chapter 19
Prompt caching
-
Figure 19.4 · Chapter 19
Streaming changes the wait you feel, not the wait itself
-
Figure 19.5 · Chapter 19
The accuracy–cost–latency triangle
-
Figure 20.2 · Chapter 20
Two shapes for serving a long-running task
-
Figure 20.3 · Chapter 20
Durable execution
-
Figure 20.4 · Chapter 20
The defensive plumbing between your traffic and the provider’s limit
-
Figure 20.5 · Chapter 20
The rollout ladder
Part VII Applications and the Road Ahead
30 diagrams
-
Figure 21.2 · Chapter 21
The coding-agent loop
-
Figure 21.3 · Chapter 21
The spectrum of intent
-
Figure 21.5 · Chapter 21
The transferable template, in one contrast
-
Figure 22.1 · Chapter 22
The canonical loop and its paperwork
-
Figure 22.2 · Chapter 22
Spec-driven development inverts the usual hierarchy
-
Figure 22.3 · Chapter 22
Test-driven development wired to the test-commit-or-revert guardrail
-
Figure 22.5 · Chapter 22
The explainer packet replaces diving into a raw diff with an ordered reading path
-
Figure 23.1 · Chapter 23
The research loop
-
Figure 23.3 · Chapter 23
Four partial substitutes for a compiler of facts
-
Figure 23.5 · Chapter 23
Where a candidate use case earns its keep
-
Figure 24.1 · Chapter 24
The target of trust calibration
-
Figure 24.2 · Chapter 24
Show your work as a curated pack, not a transcript
-
Figure 24.3 · Chapter 24
Two ways to redirect a running agent
-
Figure 25.1 · Chapter 25
The eight recipes on the autonomy slider
-
Figure 25.2 · Chapter 25
The autoresearch loop
-
Figure 25.3 · Chapter 25
The self-maintaining knowledge base
-
Figure 25.4 · Chapter 25
Deep research seen as compression
-
Figure 25.5 · Chapter 25
The ambient watcher
-
Figure 25.6 · Chapter 25
Unstructured to structured, at volume
-
Figure 25.7 · Chapter 25
The queue triage and router
-
Figure 25.8 · Chapter 25
The digital coworker
-
Figure 25.9 · Chapter 25
The premortem, or red-team simulator
-
Figure 26.2 · Chapter 26
The three roles an in-prompt example can play, placed on the class map they teach
-
Figure 26.3 · Chapter 26
The same ticket judged two ways, with the order reversed
-
Figure 26.4 · Chapter 26
The chapter’s development discipline made spatial: one pool of labeled tickets divided into three disjoint homes
-
Figure 26.5 · Chapter 26
The screen-then-detail cascade, with band widths standing in for volume
-
Figure 27.2 · Chapter 27
The generator-and-gate loop at the frontier of verification
-
Figure 27.3 · Chapter 27
The ecosystem as six durable layers with two cross-cutting rails
-
Figure 27.4 · Chapter 27
Lock-in inverted
-
Figure 27.5 · Chapter 27
One practitioner’s model of working leverage as overlapping waves, drawn as an ordering rather than a calendar—the time axis carries no dates
Appendices
1 diagram
