Agent Architecture and Design — Study Note
Personal study note covering the structural design of agentic AI systems for the NCP-AAI certification. Summarises the five core agent components (perceiver, planner, executor, memory, tool interface), three dominant control loop patterns (ReAct, plan-and-execute, reflection), and multi-agent topologies (hierarchical, peer-to-peer, role-based crews). Includes trade-offs between stateful and stateless agent design, and tool integration best practices.
Key Concepts
- Core components: perceiver · planner · executor · memory · tool interface
- ReAct loop: Observe → Reason → Act → Observe (chain-of-thought drives action selection)
- Plan-and-execute: upfront full decomposition by a planner agent; separate executors carry out steps — reduces hallucination in long-horizon tasks, sacrifices adaptability
- Reflection pattern: self-critique step post-execution before committing output
- Multi-agent topologies: hierarchical (orchestrator + workers), peer-to-peer (debate/consensus), role-based crews (CrewAI)
- Stateful vs stateless: stateless scales horizontally but requires full context per call; common hybrid uses stateless agents backed by external state store (Redis, vector DB)
- Tool integration: JSON-schema-described functions; narrow and single-purpose; rate limiting at tool layer not agent layer
Key Design Decisions
- Control loop choice (ReAct vs plan-and-execute) is the highest-leverage architectural decision — it cascades into evaluation strategy and deployment complexity
- State model choice (stateful vs stateless) determines horizontal scaling approach
- Tool layer is where error handling and rate limiting belong, keeping agent logic clean
Connections
- NCP-AAI Part 0 — covers agent principles and architecture components
- NCP-AAI Part 1 — covers LLM architecture, CrewAI framework, and LLM limitations
- Agent Development topic area — practical implementation of these architectural patterns
- Cognition, Planning, and Memory topic area — memory component detail