Chapter 3.4 — Architecture patterns

🎯 Objective

Map patterns with their when-to-use and when-to-avoid.

📊 Pattern table

Pattern	When to use	When to avoid	Risks	Controls
Single-agent	Short task, few tools	Long critical process	Tool misuse	Tool limits, evals
Planner-executor	Auditable multi-step task	Trivial task	Bad/expensive plan	Structured plan, approval
Router	Several models/pipelines	Little variation	Wrong routing	Metrics and fallback
Evaluator/Critic	Quality and safety	As the only guarantee	Judge bias	Golden set and human
Multi-agent	Real parallelism	Simple automation	Cost and coordination	Supervisor, budgets
Workflow-first	Corporate process	Open exploratory task	Rigidity	Exceptions and HITL
Event-driven	Asynchronous operation	Immediate response	Duplication	Idempotency
Agent as a service	Corporate reuse	Local prototype	Broken contract	Versioned API
Supervisor	Specialized subagents	Team without the need	Bottleneck	Policies and tracing
State machine + LLM	Critical processes	Creative exploration	Complexity	Workflow engine

🛠️ Workflow-first as a defensible pattern

For most enterprise cases, a workflow-first agent (a deterministic workflow controlling the flow, with the LLM at specific points) tends to be more defensible than free autonomous agents.

flowchart TD
    START[Start] --> C1[Classify case]
    C1 --> RAG[Retrieve evidence]
    RAG --> LLM[LLM synthesizes]
    LLM --> DEC{Critical action?}
    DEC -->|Yes| HITL[Human approval]
    DEC -->|No| ACT[Execute action]
    HITL --> ACT
    ACT --> END[End]

🚨 About multi-agent

Multi-agent multiplies cost, latency and risk. Use it only if there is:

• real decomposition;
• useful parallelism;
• necessary specialization;
• a communication protocol;
• a stopping mechanism;
• supervision.

---