Architect's Map
Agentic AI Architecture Map
Concepts, patterns, implementation and AWS mappings
24 concepts
Last reviewed · Aug 2026 ·
v0.7
All
Foundations
Runtime & Control
Tools & Integration
Context & Information
Decisions & Orchestration
Identity & Governance
Filter by pattern
A2A
Action Proposal
Agent Discovery
Agentic Workflow
Anti-Corruption Layer
Asynchronous Evaluation
Asynchronous Intake
Authorized Retrieval
Autonomy Spectrum
Bounded Loop
Canonical Task Envelope
Capability Adapter
Capability Composition
Centralized Goal Ownership
Checkpoint / Resume
Completion Gate
Context Assembly
Contract Chain
Control Plane
Controlled Agent
Curated Memory
Default Deny
Defense in Depth
Delegated Subgoal
Deterministic Sandwich
Deterministic Workflow
Distributed Tracing
Dual Principal
Durable Execution
Dynamic Replanning
End-to-End Correlation
Evidence Provenance
Federated Ownership
Governed Gateway
Human-in-the-Loop
Hybrid Orchestration
Idempotent Command
Idempotent Consumer
Information Taxonomy
Least Privilege
Least-Complexity-First
Model Gateway
Next-Action Reasoning
No-Progress Detection
Observe–Propose–Act
On-Behalf-Of
Orchestrator–Specialist
Peer Collaboration
Plan-and-Execute
Policy Enforcement
Pre / Post Gates
Pre-Context Authorization
Progress Map
Provider Abstraction
Quality Signals
Quota & Cost Governance
Relevance–Authority–Freshness
Result Normalization
Retrieval as Tool
Routing & Fallback
Semantic Tool Contract
Simplest-Pattern-First
Single-Agent Loop
State Machine
Suitability Scorecard
System of Record
Task & Artifact Exchange
Tool Scoping
Trusted Identity Context
Uncertainty–Consequence Matrix
Filter by AWS
API Gateway
API Gateway / ALB
AgentCore Gateway
AgentCore Harness
AgentCore Identity
AgentCore Memory
AgentCore Policy
AgentCore Runtime
Amazon Bedrock
Amazon Cognito
Bedrock Knowledge Bases
CloudWatch
DynamoDB
ECS / Fargate
EventBridge
IAM
Lambda
Lambda / ECS
Managed Prometheus
OpenSearch Serverless
RDS / DynamoDB
S3
SQS
Step Functions
Clear all filters
Decisions & Orchestration
Single-Agent Architecture
Use one reasoning lifecycle when a single agent can own the goal, state, actions, and completion boundary.
Single-Agent Loop
Centralized Goal Ownership
Capability Composition
AgentCore Runtime
Amazon Bedrock
Decisions & Orchestration
Peer Multi-Agent Collaboration
Coordinate independent agents as peers when authority and goal ownership are intentionally distributed.
Peer Collaboration
Federated Ownership
Task & Artifact Exchange
AgentCore Runtime
AgentCore Gateway
Decisions & Orchestration
Orchestrator–Specialist Architecture
Delegate bounded subgoals only when a specialist genuinely needs its own reasoning and goal lifecycle.
Orchestrator–Specialist
Delegated Subgoal
AgentCore Runtime
Amazon Bedrock
Tools & Integration
A2A Agent Interoperability
Standardize discovery, task exchange, status, and artifacts between independently deployed agents.
A2A
Agent Discovery
AgentCore Runtime
AgentCore Gateway
Foundations
Agentic AI & Autonomy
Understand where model-directed action adds value and where deterministic control must remain.
Autonomy Spectrum
Least-Complexity-First
Human-in-the-Loop
Amazon Bedrock
AgentCore Runtime
Foundations
Agent Suitability
Decide whether a use case needs an adaptive agent or a simpler workflow, assistant, or model call.
Suitability Scorecard
Uncertainty–Consequence Matrix
Simplest-Pattern-First
Foundations
Agentic Workflow vs Controlled Agent
Distinguish bounded model decisions inside a known path from a runtime-owned adaptive loop.
Deterministic Workflow
Agentic Workflow
Controlled Agent
Step Functions
AgentCore Runtime
Runtime & Control
Controlled Agent Architecture
Separate reasoning, lifecycle coordination, execution authority, business truth, and human accountability.
Deterministic Sandwich
Control Plane
Contract Chain
AgentCore Runtime
AgentCore Gateway
Amazon Bedrock
Runtime & Control
Execution Lifecycle & State
Track one goal execution through iterations, waiting states, checkpoints, retries, and terminal outcomes.
State Machine
Checkpoint / Resume
Durable Execution
AgentCore Runtime
DynamoDB
EventBridge
Runtime & Control
Stop, Resume & Completion
Bound the loop, detect non-progress, pause safely, and verify that the business goal is truly complete.
Bounded Loop
No-Progress Detection
Completion Gate
AgentCore Runtime
CloudWatch
SQS
Runtime & Control
Observability, Tracing & AI Evaluation
Explain system behavior and measure AI quality through correlated telemetry and independent evaluation.
Distributed Tracing
AI Evaluation
CloudWatch
Managed Prometheus
Tools & Integration
Deterministic Intake & Canonicalization
Turn varied requests and events into validated, deduplicated, durable tasks before agent reasoning begins.
Canonical Task Envelope
Idempotent Intake
SQS
DynamoDB
Tools & Integration
Action, Tool & Observation
Translate semantic model intent into governed execution and a stable meaning for the next decision.
Observe–Propose–Act
Action Proposal
Result Normalization
AgentCore Gateway
Lambda
CloudWatch
Tools & Integration
Tool Contracts & Safe Execution
Expose narrow business capabilities and execute them through pre-action and post-action control gates.
Semantic Tool Contract
Pre / Post Gates
Idempotent Command
AgentCore Gateway
Lambda
DynamoDB
Tools & Integration
Model Gateway & Provider Routing
Govern inference through provider abstraction, routing, resilience, quotas, and cost-aware policy.
Model Gateway
Routing & Fallback
Amazon Bedrock
CloudWatch
Tools & Integration
MCP & Capability Gateways
Discover and invoke AI-facing capabilities through semantic, standardized, and governed contracts.
Capability Adapter
Semantic Tool Contract
Governed Gateway
AgentCore Gateway
AgentCore Runtime
Bedrock Knowledge Bases
Context & Information
Context Engineering
Build a deliberate decision snapshot from a much larger information world available to the application.
Context Assembly
Relevance–Authority–Freshness
Tool Scoping
Amazon Bedrock
AgentCore Memory
Bedrock Knowledge Bases
Context & Information
State, Memory & Knowledge
Give execution truth, interaction history, and external reference material distinct owners and lifecycles.
Information Taxonomy
Curated Memory
System of Record
RDS / DynamoDB
AgentCore Memory
Bedrock Knowledge Bases
Context & Information
RAG in Agentic Systems
Treat retrieval as a bounded evidence capability that can influence—but not own—the next action.
Retrieval as Tool
Authorized Retrieval
Evidence Provenance
Bedrock Knowledge Bases
OpenSearch Serverless
AgentCore Gateway
Decisions & Orchestration
Reactive Agent Loop
Choose one useful next action from current state, observe the result, and rebuild context before continuing.
Next-Action Reasoning
Hybrid Orchestration
Deterministic Sandwich
AgentCore Runtime
Amazon Bedrock
AgentCore Gateway
Decisions & Orchestration
Planning & Replanning
Add explicit decomposition only when long-range dependencies make local next-action reasoning insufficient.
Plan-and-Execute
Dynamic Replanning
Progress Map
AgentCore Runtime
Amazon Bedrock
AgentCore Harness
Identity & Governance
Human vs Workload Identity
Keep the requesting person and the executing software as distinct principals with independent permissions.
Dual Principal
Least Privilege
Trusted Identity Context
IAM
AgentCore Identity
Amazon Cognito
Identity & Governance
Delegated Authority & Layered Authorization
Preserve user authority downstream and enforce permissions across runtime, capability, retrieval, and domain boundaries.
On-Behalf-Of
Defense in Depth
Pre-Context Authorization
AgentCore Identity
AgentCore Gateway
IAM
Identity & Governance
Policy, Cedar & Human Approval
Combine deterministic capability policy with durable, accountable approval for consequential actions.
Default Deny
Policy Enforcement
Human-in-the-Loop
AgentCore Policy
AgentCore Gateway
RDS / DynamoDB