Part III. Building AI Workers
- Approved by
- Panaversity ·
Note
You are here: Part III. Building AI Workers
| Journey | Where you are |
|---|---|
| Capability | You can manage an AI Worker. This part teaches you to build one. |
| Certification | Prepares you for PCDV-F, the Developer exam at the end of this part, and then for Anthropic's CCDV-F. |
| Economic | You can build and run a worker, not only direct one. With PCDV-F, the Brightline AP plugin completes portfolio milestone B. |
- Portfolio artifact: the Brightline AP plugin.
- Who reads it: builders read all seventeen chapters. Domain experts read Required and Co-build chapters in full, and the "For managers" summary of each Manager chapter. The everyday professional route and the enterprise and national leader route skip this part.
- Before you start: Part II, and a computer you can install software on.
- What comes next: PCDV-F, then Part IV, which runs and governs the worker you built, with PCAR-F after it.
What this part is about
You build your AI Worker with a coding agent. You decide what it may touch, read its plan, and check its result.
One habit runs through every chapter: before the agent builds anything, write down what you expect it to produce. That note is your own check, and you compare the agent's result against it.
Every chapter teaches Claude first, because the Developer exam is about Claude, then OpenAI, then open source. Each lab builds on Claude, then moves the same build to the other two in shorter tasks. With only one of them, you can still do every core task.
What you will be able to do
By the end of this part you can:
- build skills, a plugin and a connector with a coding agent, and review what it wrote
- run the same worker two ways: as a hosted agent, where a ready-made agent service runs it for your app, and as a self-hosted runtime, where you run it yourself
- add a harness, structured output (answers in a fixed format a program can read), subagents, context discipline (keeping the right facts in view on long work) and a cost budget (a limit on what the worker spends on model use)
- move each build from Claude to OpenAI and to open source, and record what changed
Your role in the running project: builder
In Parts I and II you described the AP Worker and managed it. Now you build its parts in brightline-ap, a TypeScript repository, which is a project folder whose history git saves. Chapters 12 to 15 set it up and work on it. From Chapter 16 it holds the Brightline AP plugin, and each later chapter adds one capability.
The seventeen chapters
Stage 1: Direct a coding agent, then read and review its code
| Chapter | What you can do after it |
|---|---|
| 12. Agentic Coding: Claude Code, Codex and OpenCode | Direct all three coding agents on a real repository |
| 13. Programming Fundamentals in the AI Era: TypeScript and Python | Read, test and change the code a coding agent wrote, in either language |
| 14. Structuring Programs in the AI Era: Objects, Data and Async | Read and review bigger programs: classes, types, schemas (rules for the shape of data) and code that waits on other systems |
| 15. Trusting Code in the AI Era: Tests, Types, Pull Requests and Code Review | Review many agent-written changes, checking risky ones more closely |
Stage 2: Build the worker's parts
| Chapter | What you can do after it |
|---|---|
| 16. Building Skills and Plugins with Coding Agents | Package professional work as skills in a plugin |
| 17. Building Connectors: Tool Design and MCP Servers | Describe the connector you need, have a coding agent build it, and put it online |
| 18. Securing Connectors: OAuth for MCP, with Better Auth and D1 | Make each person sign in, so the connector knows who is calling and for which company |
| 19. Business Data and Retrieval: Postgres, Row-Level Security and pgvector, on Neon | Keep each company's data apart, and answer from its documents with citations |
Stage 3: Run it beyond your own computer
| Chapter | What you can do after it |
|---|---|
| 20. Hosted Agents: Claude, OpenAI and Open Source | Run the worker inside your own product, with its own account and permissions |
Stage 4: Look inside and engineer it
| Chapter | What you can do after it |
|---|---|
| 21. From Requirements to Specification | Turn requirements into a clear specification, and plan how it changes over its life |
| 22. The Agent Loop, Sessions and Debugging | Write the agent loop by hand, and find bugs by reading its step-by-step log |
| 23. Harness Engineering | Limit, check and gate what the agent does |
| 24. Prompting for Reliable Structured Output | Get output a program can trust |
| 25. Agent Architecture: Workflows, Supervisors and Subagents | Choose between a fixed workflow and an agent, and split work across subagents |
| 26. Context Engineering | Keep the right facts in view across long work |
| 27. APIs and SDKs: Building Your Own Runtime | Use models in your own code and run the agent yourself |
| 28. Model Selection and Token Economics | Choose models and control cost |
How to read this part
- Start with Chapter 12, and read in order at least through Chapter 15, because each chapter starts from the repository the last one left.
- Do each lab when you reach it, and repeat it on OpenAI and open source where you can, because the plugin is built from them.
- If you get stuck, ask Zia Tutor AI. It is built on this book, and it can explain this part and quiz you on it.
- Owner
- Panaversity