Diagram Lab
Working reference for the four diagram components. Not part of the book.
Every figure below is the real content from The Agentic Engineer's Map, rendered through the components rather than as ASCII. Delete this page once the book's diagrams have been converted.
Layers
User
streams, approvals, interruptions
InterfacePart XIII
- streaming
- approvals
- provenance
- undo
intent
Agent RuntimeParts II, VII
modeldecidetoolobservedecide…
- state
- memory
- routing
- subgraphs
- stopping
what does it need to know?
Context & KnowledgeParts III–V
- prompt
- retrieval
- memory
- embeddings
- SQL
- graph
what should happen?
Durable ExecutionParts X–XII
- retries
- timers
- signals
- approval
- compensation
effects on the real world
Tools & DataParts VI, VIII, IX
- APIs
- DBs
- MCP
- SaaS
External world
Cross-cutting
Evals · Observability · Reliability · SecurityParts XIV–XVII
- none of the above is safe without them
Ladder
Who should own this state?
- The model is choosing, reasoning, or judging
- Agent Runtime
- Conversation position, intermediate results, the plan
- Graph stateLangGraph
- A business process that must survive hours, days, deploys, crashes, and retries
- WorkflowTemporal
- Facts the business considers true
- Database / APIs
Decision
The task
Is the sequence known?
yes
Does it need to survive failure?
no
Function
yes
Workflow
no
Does it decide more than once?
no
Rules
yes
Agent
Is one context enough?
yes
One agent
no
Multi-agent
Flow
- model
- decide
- tool
- observeresult enters context
until the model stops calling tools
render card
inline in the session
the review queue
the audit record