OpenBook · Forward Deployed Engineer
The engineering discipline of shipping AI applications — a book written for engineers who "get AI running on-site at the customer."
Start reading GitHub repo English version
What this book is
Three chapters on role definition and mental models, fifteen chapters on hands-on engineering you can run end-to-end in an AWS account (Part VI's three chapters cover the three scaling patterns for agents: Tool / MCP / Skill), and four appendices with reusable customer onboarding templates. The accompanying code lives in demos/ in the repo, torn down after each chapter.
demos/
，每章用完即拆。
The formula:
Outcome = Harness × Customer Harness 提供能力，Customer 提供约束。FDE 的工作是把 Harness 装到 Customer 身上。
Who this book is for
- 5+ years of engineering experience, able to independently own a backend service
- Have run at least one customer PoC, and gotten stuck at least once during a production handoff
- Have called LLM APIs and built demos, but never run an agent in a customer environment
- Can read OpenAPI specs directly, run SQL, and read distributed traces without an intermediary layer
If you've never done customer-facing work, or never called an LLM API — build a few small projects first. This book doesn't teach those.
How to read it
The most useful way is to read it alongside a real FDE project. At the end of each chapter, ask yourself:
- Does my current project have the problem this chapter describes?
- If so, how did I handle it before?
- Where does this chapter's approach differ from mine? Which is better?
A full read-through takes 6–7 hours; reading it alongside a real project over 2–3 weeks may be more useful.
Feedback
GitHub Issues is the fastest channel. What I most want:
- Counterexamples: specific scenarios where this book's methods fail in your context
- Real numbers: your PoC conversion rate, customer NPS, first-week deliverables
- Missing chapters: key engineering topics you think this book leaves out
The next edition will absorb the strongest signals.