Huang Jia/黄佳

Lead Research Engineer at A*STAR | Author of Designing AI Agents | Applied GenAI, RAG, AI Agents, and Enterprise AI Systems

[Huang Jia/黄佳 @shidaren]

Why the Harness Matters More Than the Model https://t.co/xSwbNTzze4

[Huang Jia/黄佳 @shidaren]

@zhanghaili0610 @LangChain 支持海立老师！很期待这个开源课程成为 LangChain 官方生态与中文开发者之间的一座桥梁。Deep Agents口碑非常好，国内 Agent 开发者需要高质量、工程化的学习材料，ADPS 也很愿意在内容传播、实践案例和社区交流上一起联动。

[Harry Zhang @zhanghaili0610]

A gift to the Chinese developer community for anyone building agents on @LangChain:

After 2 books on the ecosystem, I open-sourced the third — Deep Agents in Action. 8 chapters live and counting, on the Agent Harness: planning, context engineering, sub-agents, Skills, memory.

[Huang Jia/黄佳 @shidaren]

Read more:

Manning MEAP: https://t.co/SG0nIbjrko

The position paper underneath the framework (arXiv): https://t.co/5rfKHgw4Cn

cc @ManningBooks 5/5

[Huang Jia/黄佳 @shidaren]

Most agent pattern lists today sort by one axis only. That's why "Plan-and-Execute" means three different things across Ng, Anthropic, and LangChain.

A coordinate system fixes this without anyone changing vocabulary.

Not a new list. A grid the existing lists fit inside. 4/

[Huang Jia/黄佳 @shidaren]

The book proposes a two-axis framework.

Cognitive function — perception, memory, reasoning, action, reflection, collaboration, governance.

Execution topology — chain, route, parallel, orchestrate, loop, hierarchy.

7 × 6 = 42 cells. 28 are named patterns. 14 are empty. 3/

[Huang Jia/黄佳 @shidaren]

Most AI agent failures in production aren't model failures. They're design failures.

The model does its job. The system around it is what breaks: wrong scope in prompts, wrong signal in memory, wrong authority granted to tools, observability silent at the worst moments. 2/

[Huang Jia/黄佳 @shidaren]

Designing AI Agents hit #1 on Manning's Bestseller list for May 31.

Thanks to everyone who picked up the book in launch week.

A short thread on what the book argues. 1/

[Huang Jia/黄佳 @shidaren]

Check out my latest article: Your agent's eval pipeline is probably testing the wrong thing https://t.co/DYpxmNrZio via @LinkedIn

[Huang Jia/黄佳 @shidaren]

Companion resources, all open:

📄 Position paper (arXiv:2605.13850) — established the framework timestamp two months before MEAP
🧩 Framework catalog: https://t.co/nmlnUNjYmU
🛠 Book runtime (Argus): https://t.co/i3BLQ9YXtA

8/8

[Huang Jia/黄佳 @shidaren]

All of this in Designing AI Agents (@ManningBooks).

MEAP launches today. Ten chapters, twenty-eight patterns, four real-world case studies, reference appendix with the full matrix.

Sitewide 50% off through June 9 to mark the launch:

https://t.co/o7B5l14Tkh

7/

[Huang Jia/黄佳 @shidaren]

Operationalized through the Pattern Selection Card:

ASSESS what cognitive functions the task needs → ROUTE to topology that fits constraints → SELECT named patterns at those intersections

"I'm building an agent for X" becomes a candidate pattern list in 15 minutes.

6/

[Huang Jia/黄佳 @shidaren]

The framework I propose: two axes.

Cognitive function (7): Perception · Memory · Reasoning · Action · Reflection · Collaboration · Governance

Execution topology (6): Chain · Route · Parallel · Orchestrate · Loop · Hierarchy

7×6 grid. 28 named patterns at the intersections.

5/

[Huang Jia/黄佳 @shidaren]

Most "agent design" today picks a topology and assumes the cognitive function is obvious. It usually isn't.

The result: copies a template, adds memory, adds a planner, adds a critic, works in demos, fails in production.

The cause isn't the framework.

4/

[Huang Jia/黄佳 @shidaren]

Different failure modes. Different testing strategies. Different cost curves. Same diagram.

A pattern is not a topology. A pattern is not a cognitive function. A pattern is the intersection.

3/

[Huang Jia/黄佳 @shidaren]

The same Orchestrator-Workers topology can mean:

→ Plan-and-Execute (planner decomposes, executors run) → Hierarchical Delegation (manager assigns specialists) → Adversarial Verification (generator + critic iterate)

Three different systems. Same boxes-and-arrows.

2/

[Huang Jia/黄佳 @shidaren]

There are two ways to describe an agent system.

Execution diagram — boxes, arrows, who calls whom. Cognitive description — what the agent perceives, remembers, decides, acts on.

Most write-ups pick one axis and stop. Neither alone identifies the agent.

1/

[Huang Jia/黄佳 @shidaren]

There are two ways to describe an agent system.

Execution diagram — boxes, arrows, who calls whom. Cognitive description — what the agent perceives, remembers, decides, acts on.

Most write-ups pick one axis and stop. Neither alone identifies the agent.

[Huang Jia/黄佳 @shidaren]

I just bought tickets to Level Up Your Career! http://t.co/KBCiH3vuf1