Architecture Selection Guide

Use these decision frameworks to pick the right architecture. Each flowchart covers one Part of this guide.

---

Master Decision: Where to Start

flowchart LR
    Q1[Start here] --> Q2{How many agents?}
    Q2 -->|One| S1["Part I: Core Patterns<br/>(1-10)"]
    Q2 -->|Multiple| Q3{Need formal<br/>collaboration?}
    Q3 -->|Yes| S2["Part IV: Collaboration<br/>(19-26)"]
    Q3 -->|No, just infrastructure| S3["Part XII: Infrastructure<br/>(59-62)"]

    S1 --> Q4{Need quality gates?}
    Q4 -->|Yes| S4["Part II: Quality<br/>(11-13)"]
    Q4 -->|No| Q5{Need parallelism?}
    Q5 -->|Yes| S5["Part III: Pipelines<br/>(14-18)"]
    Q5 -->|No| Q6{Need memory?}
    S4 --> Q6
    S5 --> Q6
    Q6 -->|Yes| S6["Part X: Memory<br/>(46-52)"]
    Q6 -->|No| Q7{Need safety?}
    S6 --> Q7
    Q7 -->|Always yes in prod| S7["Part XI: Safety<br/>(53-58)"]
    S7 --> DONE[Ship it!]

    style Q1 fill:#4CAF50,color:#fff
    style DONE fill:#4CAF50,color:#fff
    style S1 fill:#2196F3,color:#fff
    style S2 fill:#9C27B0,color:#fff
    style S3 fill:#009688,color:#fff
    style S4 fill:#FF9800,color:#fff
    style S5 fill:#FF9800,color:#fff
    style S6 fill:#3F51B5,color:#fff
    style S7 fill:#F44336,color:#fff

---

1. Core Patterns (1–10): "What kind of agent system do I need?"

flowchart TD
    START[What's your task?] --> SINGLE{Single agent<br/>enough?}

    SINGLE -->|Yes| TOOLS{Needs external<br/>tools/APIs?}
    TOOLS -->|No| STEPS{Multi-step<br/>pipeline?}
    STEPS -->|Yes| PC[2. Prompt Chaining]
    STEPS -->|No| SIMPLE[Just use a plain LLM call]

    TOOLS -->|Yes| IMPROVE{Needs to improve<br/>its own output?}
    IMPROVE -->|No| REACT[1. ReAct]
    IMPROVE -->|Yes| CRITERIA{Has quantifiable<br/>scoring criteria?}
    CRITERIA -->|Yes| EO[8. Evaluator-Optimizer]
    CRITERIA -->|No| REFL[7. Reflection]

    SINGLE -->|No, need multiple agents| CLASSIFY{How do agents<br/>divide work?}

    CLASSIFY -->|Route input to the<br/>right specialist| ROUTE[3. Routing]
    CLASSIFY -->|Each processes<br/>independently| PAR[4. Parallelization]

    CLASSIFY -->|Agents coordinate<br/>together| CONTROL{Central<br/>controller?}

    CONTROL -->|Yes, one boss| PREDEFINED{Predefined agent<br/>assignments?}
    PREDEFINED -->|Yes| SUP[6. Supervisor]
    PREDEFINED -->|No, dynamic delegation| OW[5. Orchestrator-Worker]

    CONTROL -->|No, peer-to-peer| HANDOFF{Conversational<br/>handoffs?}
    HANDOFF -->|Yes| SWARM[9. Swarm]
    HANDOFF -->|No, nested teams| HIER[10. Hierarchical Teams]

    style START fill:#4CAF50,color:#fff
    style SIMPLE fill:#9E9E9E,color:#fff
    style PC fill:#2196F3,color:#fff
    style REACT fill:#2196F3,color:#fff
    style EO fill:#2196F3,color:#fff
    style REFL fill:#2196F3,color:#fff
    style ROUTE fill:#2196F3,color:#fff
    style PAR fill:#2196F3,color:#fff
    style OW fill:#2196F3,color:#fff
    style SUP fill:#2196F3,color:#fff
    style SWARM fill:#2196F3,color:#fff
    style HIER fill:#2196F3,color:#fff

---

2. Quality & Verification (11–13): "How should output be verified?"

flowchart TD
    START[Need output verification?] --> TYPE{Verification type?}

    TYPE -->|Objective: run code,<br/>check tests, pass/fail| GV[11. Generator-Verifier]
    TYPE -->|Subjective: LLM scores<br/>against rubric| JUDGE[12. Agent-as-a-Judge]
    TYPE -->|Human approval<br/>required| HITL[13. Human-in-the-Loop]

    style START fill:#4CAF50,color:#fff
    style GV fill:#FF9800,color:#fff
    style JUDGE fill:#FF9800,color:#fff
    style HITL fill:#FF9800,color:#fff

Already covered in Part I: #7 Reflection (self-critique) and #8 Evaluator-Optimizer (scored iteration) are also quality patterns. Use Part II when you need external verification.

---

3. Parallelism & Pipelines (14–18): "How should work be distributed?"

flowchart TD
    START[Need parallel or<br/>pipeline execution?] --> INPUT{What varies?}

    INPUT -->|Large input needs<br/>chunking| MR[14. Map-Reduce]
    INPUT -->|Want model diversity<br/>for quality| MOA[15. Mixture-of-Agents]
    INPUT -->|Tasks have complex<br/>dependencies| DAG[16. DAG Orchestration]
    INPUT -->|Need explicit plan<br/>before execution| PE[17. Plan-and-Execute]
    INPUT -->|Want to minimize<br/>cost per query| CASC[18. Cascading Agents]

    style START fill:#4CAF50,color:#fff
    style MR fill:#FF9800,color:#fff
    style MOA fill:#FF9800,color:#fff
    style DAG fill:#FF9800,color:#fff
    style PE fill:#FF9800,color:#fff
    style CASC fill:#FF9800,color:#fff

---

4. Multi-Agent Collaboration (19–26): "How should agents work together?"

flowchart TD
    START[How should agents<br/>collaborate?] --> STRUCT{Communication<br/>structure?}

    STRUCT -->|Shared workspace,<br/>no direct communication| BB[19. Blackboard]
    STRUCT -->|Competitive bidding| BID{Formal protocol?}
    BID -->|Informal, market-style| MKT[20. Market-Based]
    BID -->|Formal FIPA protocol| CNP[21. Contract Net]

    STRUCT -->|Debate & argue| PURPOSE{Goal?}
    PURPOSE -->|Improve accuracy via<br/>disagreement| DEBATE[22. Debate]
    PURPOSE -->|Find vulnerabilities| RED[23. Red-Team]

    STRUCT -->|Structured roles &<br/>pipelines| FRAMEWORK{Preferred style?}
    FRAMEWORK -->|Fixed roles, persona-driven| CREW[24. Role-Based / CrewAI]
    FRAMEWORK -->|Open discussion, multi-turn| AUTOGEN[25. Conversational / AutoGen]

    STRUCT -->|Async event-driven| EVENT[26. Event-Driven]

    style START fill:#4CAF50,color:#fff
    style BB fill:#9C27B0,color:#fff
    style MKT fill:#9C27B0,color:#fff
    style CNP fill:#9C27B0,color:#fff
    style DEBATE fill:#9C27B0,color:#fff
    style RED fill:#9C27B0,color:#fff
    style CREW fill:#9C27B0,color:#fff
    style AUTOGEN fill:#9C27B0,color:#fff
    style EVENT fill:#9C27B0,color:#fff

---

5. Reasoning Patterns (27–30): "How should the agent think?"

flowchart TD
    START[How should the<br/>agent reason?] --> VIS{Should reasoning<br/>be visible to user?}

    VIS -->|No, hide internal thinking| IM[27. Inner Monologue]
    VIS -->|Yes / doesn't matter| SPEED{Latency critical?}
    SPEED -->|Yes, speculate on<br/>multiple paths| SE[28. Speculative Execution]
    SPEED -->|No| OUTPUT{Output type?}
    OUTPUT -->|Long structured document| SOT[29. Skeleton of Thought]
    OUTPUT -->|Problem needing<br/>multi-strategy exploration| REACTREE[30. ReAcTree]

    style START fill:#4CAF50,color:#fff
    style IM fill:#795548,color:#fff
    style SE fill:#795548,color:#fff
    style SOT fill:#795548,color:#fff
    style REACTREE fill:#795548,color:#fff

---

6. Domain Applications (31–34): "What domain am I working in?"

flowchart TD
    START[Domain-specific<br/>agent needed?] --> DOMAIN{What domain?}

    DOMAIN -->|Document Q&A,<br/>knowledge retrieval| ARAG[31. Agentic RAG]
    DOMAIN -->|Write, test,<br/>debug code| ACODE[32. Agentic Coding]
    DOMAIN -->|Need new tools<br/>at runtime| TTYPE{How?}
    TTYPE -->|Ad-hoc, on-the-fly| STA[33. Self-Tooling Agent]
    TTYPE -->|Systematic pipeline<br/>with testing| TF[34. ToolFactory]

    style START fill:#4CAF50,color:#fff
    style ARAG fill:#00BCD4,color:#fff
    style ACODE fill:#00BCD4,color:#fff
    style STA fill:#00BCD4,color:#fff
    style TF fill:#00BCD4,color:#fff

---

7–8. Cognitive Architectures & Advanced Search (35–41)

flowchart TD
    START[Need advanced<br/>reasoning?] --> TYPE{What kind?}

    TYPE -->|Combine LLM with<br/>formal logic/rules| PARADIGM{Always use both?}
    PARADIGM -->|Yes, both together| NS[35. Neuro-Symbolic]
    PARADIGM -->|Route: simple→neural,<br/>complex→symbolic| DP[36. Dual-Paradigm]

    TYPE -->|Systematic search<br/>over solution space| COMPLEXITY{Problem complexity?}
    COMPLEXITY -->|Very high, need<br/>backtracking| MCTS{Need reflection?}
    MCTS -->|No, standard MCTS| LATS[37. LATS]
    MCTS -->|Yes, per-simulation| IMCTS[38. I-MCTS]
    MCTS -->|Yes, cross-episode| RMCTS[39. R-MCTS]
    COMPLEXITY -->|Multi-agent search| COTS[40. CoTS]
    COMPLEXITY -->|Moderate, top-K paths| BEAM[41. Beam Search]

    style START fill:#4CAF50,color:#fff
    style NS fill:#607D8B,color:#fff
    style DP fill:#607D8B,color:#fff
    style LATS fill:#673AB7,color:#fff
    style IMCTS fill:#673AB7,color:#fff
    style RMCTS fill:#673AB7,color:#fff
    style COTS fill:#673AB7,color:#fff
    style BEAM fill:#673AB7,color:#fff

---

9. Agent Protocols (42–45): "How should agents connect?"

flowchart TD
    START[What connectivity<br/>do you need?] --> TYPE{Connecting to what?}

    TYPE -->|Tools, databases, APIs<br/>from a single agent| MCP[42. MCP]
    TYPE -->|Other agents from<br/>different vendors| SCALE{Scale?}
    SCALE -->|Enterprise,<br/>Google ecosystem| A2A[43. A2A Protocol]
    SCALE -->|Lightweight, REST-based,<br/>framework-agnostic| ACP[44. ACP]
    SCALE -->|Internet-scale, decentralized<br/>open marketplace| ANP[45. ANP]

    style START fill:#4CAF50,color:#fff
    style MCP fill:#E91E63,color:#fff
    style A2A fill:#E91E63,color:#fff
    style ACP fill:#E91E63,color:#fff
    style ANP fill:#E91E63,color:#fff

Protocol	Created By	Transport	Best For	Maturity
MCP	Anthropic / Linux Foundation	JSON-RPC over stdio/SSE	Agent-to-tool	Production-ready
A2A	Google	HTTP + JSON-RPC	Agent-to-agent (enterprise)	Early adoption
ACP	IBM BeeAI	REST (HTTP verbs)	Agent-to-agent (lightweight)	Experimental
ANP	Community	DID + HTTP	Agent-to-agent (decentralized)	Research

---

10. Agent Memory (46–52): "What should the agent remember?"

flowchart TD
    START[What does the agent<br/>need to remember?] --> WHAT{Type of information?}

    WHAT -->|Specific past events<br/>with timestamps| EP[46. Episodic Memory]
    WHAT -->|Facts, concepts,<br/>relationships| SEM[47. Semantic Memory]
    WHAT -->|How to do things<br/>— learned routines| PROC[48. Procedural Memory]
    WHAT -->|Interconnected notes<br/>with dynamic links| AMEM[49. Agentic Memory]
    WHAT -->|Shared knowledge across<br/>multiple agents| COLLAB[50. Collaborative Memory]
    WHAT -->|Past successful<br/>task trajectories| CER[51. CER]
    WHAT -->|Deciding what to<br/>remember or forget| META[52. Meta-Memory]

    EP -.->|Feeds into| SEM
    SEM -.->|Informs| PROC
    META -.->|Manages all| EP
    META -.->|Manages all| SEM

    style START fill:#4CAF50,color:#fff
    style EP fill:#3F51B5,color:#fff
    style SEM fill:#3F51B5,color:#fff
    style PROC fill:#3F51B5,color:#fff
    style AMEM fill:#3F51B5,color:#fff
    style COLLAB fill:#3F51B5,color:#fff
    style CER fill:#3F51B5,color:#fff
    style META fill:#3F51B5,color:#fff

Tip: Real agents typically combine 2–3 memory types. Episodic + Semantic + Procedural mirrors the human memory system. Add Meta-Memory when scale requires selective forgetting.

---

11. Agent Safety (53–58): "What can go wrong and how do I prevent it?"

flowchart TD
    START[What failure mode<br/>are you defending against?] --> THREAT{Threat type?}

    THREAT -->|Harmful/toxic content| GA[53. Guardrail Agent]
    THREAT -->|External service outages| CB[54. Circuit Breaker]
    THREAT -->|Multi-step task fails midway| SAGA[55. Saga Pattern]
    THREAT -->|Unauthorized access| ACCESS{Scope?}
    ACCESS -->|Centralized validation| GW[56. API Gateway]
    ACCESS -->|Per-task scoped credentials| EI[57. Ephemeral Identity]
    THREAT -->|Tasks failing silently| DL[58. Dead Letter]

    style START fill:#4CAF50,color:#fff
    style GA fill:#F44336,color:#fff
    style CB fill:#F44336,color:#fff
    style SAGA fill:#F44336,color:#fff
    style GW fill:#F44336,color:#fff
    style EI fill:#F44336,color:#fff
    style DL fill:#F44336,color:#fff

Production checklist: At minimum, every production agent needs: Guardrail (content) + Circuit Breaker (availability) + Dead Letter (observability). Add Saga for multi-step workflows, Gateway + Ephemeral Identity for multi-tenant systems.

---

Quick Reference Table

#	Architecture	Part	Complexity	Agents	Best For
1	ReAct	I	Low	1	Simple tool-augmented tasks
2	Prompt Chaining	I	Low	1	Sequential multi-step pipelines
3	Routing	I	Low-Med	1+N	Multi-domain classification
4	Parallelization	I	Medium	N	Independent concurrent subtasks
5	Orchestrator-Worker	I	Med-High	1+N	Unpredictable task decomposition
6	Supervisor	I	Med-High	1+N	Coordinated multi-agent control
7	Reflection	I	Medium	1-2	Self-improving output quality
8	Evaluator-Optimizer	I	Medium	2	Scored quality iteration
9	Swarm	I	High	N	Conversational multi-domain handoffs
10	Hierarchical Teams	I	High	N	Large-scale enterprise workflows
11	Generator-Verifier	II	Medium	2	Execution-based verification (code, math)
12	Agent-as-a-Judge	II	Medium	2	One-shot quality evaluation & ranking
13	Human-in-the-Loop	II	Medium	1+	High-stakes approval workflows
14	Map-Reduce	III	Medium	N	Processing large documents/datasets
15	Mixture-of-Agents	III	High	N models	Maximum quality via model diversity
16	DAG / Graph Orchestration	III	Medium	N	Complex dependency workflows
17	Plan-and-Execute	III	Medium	1	Strategic multi-step tasks with replanning
18	Cascading Agents	III	Medium	2-3	Cost-optimized inference
19	Blackboard	IV	High	N	Incremental collaborative problem-solving
20	Market-Based / Bidding	IV	Medium	N	Dynamic self-organizing task allocation
21	Contract Net Protocol	IV	Medium	N	Formal task delegation with accountability
22	Debate / Adversarial	IV	Medium	2-3	Rigorous decision analysis
23	Red-Team Agent	IV	Medium	2	Security & safety testing
24	Role-Based (CrewAI)	IV	Medium	N	Persona-driven team pipelines
25	Conversational (AutoGen)	IV	Medium	N	Emergent group problem-solving
26	Event-Driven Multi-Agent	IV	High	N	Reactive async multi-agent systems
27	Inner Monologue	V	Low	1	Clean output with hidden reasoning
28	Speculative Execution	V	Medium	1	Low-latency branching decisions
29	Skeleton of Thought	V	Medium	1	Fast long-form generation
30	ReAcTree	V	High	1	Multi-strategy tree exploration
31	Agentic RAG	VI	Medium	1	Intelligent document retrieval
32	Agentic Coding	VI	Medium	1	Self-healing code generation
33	Self-Tooling Agent	VI	Medium	1	Runtime tool creation
34	ToolFactory	VI	Medium	1	Systematic tool generation pipeline
35	Neuro-Symbolic	VII	High	1	LLM + formal logic/rules
36	Dual-Paradigm	VII	Medium	1	System 1/System 2 routing
37	LATS	VIII	High	1	LLM + Monte Carlo Tree Search
38	I-MCTS	VIII	High	1	Per-simulation introspective search
39	R-MCTS	VIII	High	1	Cross-episode reflective search
40	CoTS	VIII	High	N	Multi-agent collaborative tree search
41	Beam Search	VIII	Medium	1	Top-K solution exploration
42	MCP	IX	Low	—	Tool integration standard
43	A2A	IX	Medium	—	Cross-vendor agent communication
44	ACP	IX	Medium	—	Framework-agnostic agent messaging
45	ANP	IX	High	—	Internet-scale agent networking
46	Episodic Memory	X	Medium	1+	Learning from past experiences
47	Semantic Memory	X	Medium	1+	Storing facts and relationships
48	Procedural Memory	X	Medium	1+	Remembering how to do tasks
49	Agentic Memory (A-MEM)	X	High	1+	Interconnected knowledge notes
50	Collaborative Memory	X	Medium	N	Shared team knowledge
51	CER	X	Medium	1+	Prioritized experience replay
52	Meta-Memory	X	High	1+	Memory system optimization
53	Guardrail Agent	XI	Low	1+	Content safety filtering
54	Circuit Breaker	XI	Low	1+	Preventing cascading failures
55	Saga Pattern	XI	Medium	1	Multi-step rollback
56	API Gateway	XI	Low	N	Centralized access control
57	Ephemeral Identity	XI	Medium	N	Least-privilege agent execution
58	Dead Letter	XI	Low	1+	Handling failed tasks
59	Agentic Mesh	XII	High	N	Enterprise multi-agent infrastructure
60	Agent Registry	XII	Medium	N	Large-scale agent management
61	Agent Control Plane	XII	High	N	Agent governance and lifecycle
62	Data Flywheel	XII	Medium	1+	Continuous improvement from usage
63	VLA Models	XIII	Very High	1	Vision + language + robot actions
64	Embodied AI	XIII	Very High	1	Physical world agents
65	World Models	XIII	Very High	1	Internal environment simulation
66	CoALA	XIV	—	—	Agent architecture taxonomy
67	AEGIS Framework	XIV	High	1+	Comprehensive layered safety
68	ADAS	XIV	High	Meta	Automated agent design search
69	MASE	XIV	High	1+	Self-evolving agents
70	Nested Learning	XIV	High	1	Multi-timescale optimization