Five-Layer Architecture
A layered architecture model for agentic systems that separates concerns and enables scalable, maintainable, production-ready systems.
The Five Layers
code
┌─────────────────────────────────────┐ │ Layer 5: Governance │ Safety, compliance, audit ├─────────────────────────────────────┤ │ Layer 4: Runtime │ Execution, monitoring, scaling ├─────────────────────────────────────┤ │ Layer 3: Interaction │ Tools, APIs, communication ├─────────────────────────────────────┤ │ Layer 2: Context │ Memory, state, knowledge ├─────────────────────────────────────┤ │ Layer 1: Cognition │ Reasoning, planning, decisions └─────────────────────────────────────┘
Layer 1: Cognition
Core intelligence - reasoning, planning, and decision-making.
- •Task understanding and decomposition
- •Planning and strategy formulation
- •Reasoning patterns (ReAct, Chain-of-Thought)
- •Meta-cognition and self-reflection
- •Anti-pattern: Mixing business logic into prompts, no separation between reasoning and action
Layer 2: Context
Memory, state, and knowledge management.
- •Short-term working memory (conversation history)
- •Long-term memory (persistent knowledge, vector search)
- •State management and checkpointing
- •Context window optimization and compression
- •Anti-pattern: Passing entire history to every agent (context bloat)
Layer 3: Interaction
Tools, APIs, and communication channels.
- •Tool/function registry and execution
- •External API integration with error handling
- •Multi-agent communication protocols
- •Rate limiting, retry logic, serialization
- •Anti-pattern: Tight coupling, no retry/timeout, missing input validation
Layer 4: Runtime
Execution environment, monitoring, and scaling.
- •Agent lifecycle management (start, stop, restart)
- •Concurrent execution and orchestration
- •Observability (logging, tracing, metrics)
- •Resource quotas, health checks, circuit breakers
- •Anti-pattern: No observability, no resource limits, synchronous bottlenecks
Layer 5: Governance
Safety, compliance, and audit.
- •Safety guardrails and action constraints
- •Human-in-the-loop approval gates
- •Audit logging, traceability, compliance
- •PII detection, access control, rollback
- •Anti-pattern: No approval gates, missing audit trails, no rollback
Layer Interaction Patterns
Bottom-Up Flow: Cognition decides → Context provides info → Interaction executes → Runtime monitors → Governance validates
Top-Down Constraints: Governance sets policies → Runtime enforces limits → Interaction restricts tools → Context filters sensitive data → Cognition operates within constraints
Cross-Layer: Observability, security, error handling, and performance optimization span all layers.
Designing a New System
- •Start with Cognition: What reasoning/planning capability do you need?
- •Add Context: What information must agents remember?
- •Define Interaction: What tools/APIs do agents need?
- •Build Runtime: How will you execute, monitor, scale?
- •Implement Governance: What safety/compliance is required?
Quick Reference
| Layer | Focus | Key Question | Example Component |
|---|---|---|---|
| 1. Cognition | Intelligence | "What should I do?" | LLM reasoning |
| 2. Context | Memory | "What do I know?" | Vector database |
| 3. Interaction | Tools | "How do I act?" | API clients |
| 4. Runtime | Execution | "Is it working?" | Monitoring |
| 5. Governance | Safety | "Is it safe?" | Approval gates |
When to Use
- •"How should I architect my agentic system?"
- •"Where does [component] belong in the architecture?"
- •"My system is getting messy and hard to maintain"
- •"How do I separate concerns in my agent system?"
References
- •
refs/layers-deep-dive.md- Detailed responsibilities, components, and anti-patterns per layer - •
refs/implementation-guide.md- Step-by-step guide to building each layer