Huang Jia
AI Researcher at the Institute of High Performance Computing, Agency for Science, Technology and Research (A*STAR) / Author of Agent Design Patterns
Huang Jia is an AI researcher at a Singapore-based research institute, specializing in large language models, agent systems, and production-grade AI engineering. A technical author and systems-oriented thinker, he focuses on transforming complex AI capabilities into architectures that are designable, reusable, and deliverable in real-world environments. He is the author of several technical books, including Hands-on AI Agent Development with Large Models, RAG in Practice: Building Applications with Large Models, and GPT Illustrated: How Large Models Are Built. Currently, his research centers on Agent design patterns, MCP/A2A protocols, and the engineering of Agent Skills, with the goal of advancing AI from “capability demonstrations” to fully systematized engineering practices. His open-source course, Claude Code Engineering in Practice, provides a systematic exploration of four core extension mechanisms in agent frameworks (Sub-Agents, Skills, Hooks, and MCP). Through in-depth source code analysis of projects such as OpenClaw, he has proposed a comprehensive framework of 21 Agent design patterns spanning six cognitive dimensions.
Topic
Agent Design Patterns: From Cognitive Architectures to Engineering Implementation
Agent is evolving from “chat-capable AI” to “autonomous digital employees.” However, most Agent projects remain in the early stage of prompt concatenation and tool invocation, lacking systematic architectural thinking. This talk, based on the book *Agent Design Patterns*, presents a design pattern system covering six cognitive dimensions: perception, memory, reasoning, action, reflection, and collaboration. Using Claude Code (Anthropic’s official Agent) and OpenClaw (cross-platform Agent) as case studies, we will demonstrate how these patterns are orchestrated and validated in real products. The key insight: an Agent’s reliability comes not from bigger models, but from better architecture—the patterns remain, while scenarios evolve. Outline: 1. Opening: Why Agents Need Design Patterns (5 minutes) 1.1 From Software 1.0 to 3.0: three paradigm shifts—code → weights → intent 1.2 Three common anti-patterns in current Agent development: all-purpose prompts, no memory, no guardrails 1.3 Value of design patterns: reusable architectural decisions and cross-platform thinking tools 2. Six-Dimensional Cognitive Architecture: Overview of Agent Design Patterns (10 minutes) 2.1 Perception: attention focus → philosophy of selective ignoring 2.2 Memory: hierarchical memory → RAG → episodic memory 2.3 Reasoning: chain-of-thought → thought tree → thought graph 2.4 Action: ReAct → plan-execute → tool orchestration 2.5 Reflection: self-correction → reflective memory → meta-learning 2.6 Collaboration: debate → delegation → routing → collective intelligence 3. Dual Case Study: Same Patterns, Different Implementations (15 minutes) 3.1 Case 1: Claude Code — Agent architecture in developer terminal - CLAUDE.md implementing hierarchical memory - Sub-Agent implementing delegation pattern - permission_mode implementing self-correction 3.2 Case 2: OpenClaw — Agent architecture in cross-platform personal assistant - Gateway star topology implementing routing - Dual-layer memory (log layer + persistent layer) for hierarchical memory - Binding declarative routing vs. Claude Code implicit routing 3.3 Migration Table: Complete mapping of 10 patterns across both systems 4. Practical Guidelines: Five Rules for Building Reliable Agents (8 minutes) 4.1 Rule 1: Design memory before action 4.2 Rule 2: Use guardrails instead of hope 4.3 Rule 3: Match isolation granularity to trust boundaries 4.4 Rule 4: Development driven by standards (SDD) 4.5 Rule 5: Principle of minimal complexity 5. Closing: The New Human-Agent Contract (2 minutes) 5.1 Humans are responsible for intent; Agents for action 5.2 Complementary, not substitutive — from tool user to architecture designer 5.3 Power of design patterns: a thinking tool that transcends specific implementations
