Design the Agent as a product, ship it as a system
I'm an AI product manager writing in the open. These are my Agent engineering notes — from product capability and engineering foundations to harnesses, design patterns, and context engineering. Problem-driven, always growing.
If the course helped, scan with WeChat to tip a coffee. A note makes my day.
AAgent Product & Capability Design6 posts
Why Understanding Agents Starts With Splitting Them Into Five Layers
An Agent's capabilities and failures all grow on five foundational layers — Loop / Tool / Planning / Memory / Multi-Agent.
A · 1 Product · Post 1Why an Agent's First Move Shouldn't Be to Start Doing
The task completion path is an Agent's ability to break down and drive a request from need to output. PMs should design it to be observable and instrumentable.
A · 2 Product · Post 2Why You Can't Blame Agent Failures in the Abstract — Break Them Down by Node
Split Agent failures into four node types — input / tool / reasoning / output — locate first, then improve.
BAgent Engineering Foundations6 posts
Why a Stronger Model Won't Save Your Agent — Fix the Harness First
The model is the driver, the Harness is the car. If it can't deliver, it's usually the car's fault, not the driver's.
B · 1 Foundations · Post 1Why a Mismanaged Context Window Makes Even the Smartest Agent Dumb
Context is an Agent's scarcest resource—treat it as a system to be engineered, using a seven-dimension lens and a composition audit.
B · 2 Foundations · Post 2Why the First Thing to Break in Production Isn't the Model — It's the Plumbing
The Agent Gateway — ingress, routing, persistence, reliable delivery, and concurrency that let an Agent truly live in production.
CHarness Engineering8 posts
Why Capable Agents Still Fail
Most failures aren't in the model weights, but in the layer of engineering infrastructure outside the weights — the thing we call the harness.
C · 1 Harness · Post 1Same Model, Wildly Different Results — The Harness Is the Real Variable
The Harness is all of the engineering infrastructure outside the model weights — the Five Subsystems are its components, the Four Layers of Defense are how you ship it.
C · 2 Harness · Post 2Why the Repo Must Be the Agent's Single Source of Truth
Information that isn't in the repo effectively doesn't exist for an Agent. Knowledge must be committed to files, not scattered across tools.
DAgent Design Patterns (21)5 posts
A Panoramic Map of the 21 Agent Design Patterns
Lay out all 21 patterns from Google's "Agentic Design Patterns" at once, group them into 6 clusters, and mark which ones this course has already covered.
D · 1 Patterns · Post 1"Reflection and Reasoning: Letting an Agent Go Back and Fix Its Work, and Think More Deeply"
A one-pass output can be wrong without the Agent knowing it; Reflection adds critique-driven, targeted revision; Reasoning makes the thinking explicit so the Agent thinks more deeply.
D · 2 Patterns · Post 2"Learning & Adaptation + RAG: So Your Agent Stops Starting From Scratch — and Stops Making Things Up"
Learning lets an Agent improve its strategy from feedback; RAG lets it plug into your private knowledge — both work at the context layer.
EContext Engineering7 posts
What the model actually sees each turn was never the prompt you wrote
The context the model receives each turn is one big blob the framework assembled for you; only after all seven layers are 100% visible can you talk about optimizing.
E · 1 Context · Post 1Why the prompt file you're reading is often not the system the model actually receives
A single system message is usually stitched together from several sources; whether a rule belongs in a file or in code comes down to three tests—scope, priority, and robustness.
E · 2 Context · Post 2Why this turn's output shouldn't have to "both speak and be parsed"
Whether a chunk of output should speak plain language or spit out JSON has one and only one criterion—"who consumes it"; whether a tool should be eager or deferred comes down to the latency red line times the number of tools.
EXCase Studies3 studies
"Teardown: Claude Code — A Coding Agent That Nails All Five Layers"
Using the course framework as a lens to dissect Claude Code's five-layer foundation, transparency, error recovery, and context management.
EX · 2 Cursor"Teardown: Cursor — Context Engineering and Human-in-the-Loop Transparency"
Use the five-layer technical foundation and the Harness framework to take Cursor apart — see how it feeds an entire codebase into a finite window and guards trust with diffs.
EX · 3 Parrot / Lassie"Vertical AI-OS Teardown: Why Parrot and Lassie Dig One Deep Well First"
A side-by-side look at auto-repair startup Parrot and dental startup Lassie to define the "vertical deep-well Agent product" archetype.
LABLab11 projects
harness-kit
The engineering method for "running long tasks with Claude Code", distilled into a project skeleton you can clone and use.
LAB · 2 Open Sourceagent-memory-kit
A runtime memory layer for the product Agent you're building — four roles that distill experience into reusable lessons.
LAB · 3 Open Sourcecontext-engineering-kit
One CONTEXT.md to manage everything that enters the context window across seven layers — each with a budget and a caching strategy.