This tutorial keeps 24 chapters on purpose. The chapters are not separate products; each one isolates one harness mechanism so a reader can build, run, and understand it before the next layer appears. The machine-checkable version of that ladder lives in Progression Contract.
Not now. Some topics are adjacent, but merging them would make the learning curve worse:
| Adjacent Chapters | Why Keep Separate |
|---|---|
| s05 Electron Shell / s06 Sidecar / s07 Session | UI process isolation, sidecar control plane, and session lifecycle are three different failure domains. |
| s10 Workspace Memory / s11 User Memory / s12 Cloud Memory | The storage location, ownership, update policy, and retrieval timing are different. |
| s16 Skills / s17 MCP / s18 Experts | Skills add instructions, MCP adds external tools, Experts package persona + memory + tools. |
| s21 SQLite / s22 Automation / s23 Audit | Persistence, scheduling, and security governance share infrastructure but answer different production questions. |
The right improvement is enrichment, not compression: better prerequisites, clearer architecture notes, and stronger cross-chapter links.
| Chapter | Mechanism | Build Outcome |
|---|---|---|
| s01 Agent Loop | ReAct-style loop with tool calls | A minimal agent loop. |
| s02 Tool Dispatch | Tool registry and handler dispatch | Add tools without changing the loop. |
| s03 Deferred Loading | Tool discovery before schema loading | Keep startup prompt small. |
| s04 Permission & Hooks | Tiered trust and hook gates | Let the agent act inside safe boundaries. |
| Chapter | Mechanism | Build Outcome |
|---|---|---|
| s05 Electron Shell | Main/renderer/preload process split | Understand desktop app boundaries. |
| s06 Sidecar Server | JSON-RPC control plane and ring buffer | Keep agent work outside the UI process. |
| s07 Session Management | Per-session process lifecycle | Run isolated work sessions. |
| s08 Model Routing | Cheap model for routing, stronger model for reasoning | Match model cost to task complexity. |
| s09 JSONL Transcript | Append-only event stream | Recover and replay sessions. |
| Chapter | Mechanism | Build Outcome |
|---|---|---|
| s10 Workspace Memory | Project-local logs and curated memory | Remember current project work. |
| s11 User Memory | Cross-project preferences and identity | Carry stable user context. |
| s12 Cloud Memory | Server-side profile and history recall | Retrieve long-horizon context. |
| s13 Output Externalization | Swap large tool output to disk | Keep context small. |
| s14 Context Compact | Structured compression before overflow | Continue long sessions. |
| s15 Prompt Assembly | Runtime prompt composition | Assemble exactly the context needed. |
| Chapter | Mechanism | Build Outcome |
|---|---|---|
| s16 Skills System | Instruction packages with optional scripts/assets | Add capabilities without changing core code. |
| s17 MCP Connectors | External tool protocol and trust model | Connect third-party services safely. |
| s18 Experts System | Domain bundles of persona, memory, and tools | Switch working modes cleanly. |
| s19 Visualizer | Structured output to diagrams/UI artifacts | Turn agent output into inspectable views. |
| s20 Result Presentation | Prioritized deliverables | Present work as files, summaries, and diffs. |
| Chapter | Mechanism | Build Outcome |
|---|---|---|
| s21 SQLite Database | Durable metadata and usage tracking | Persist sessions and usage. |
| s22 Automation Scheduler | Timed and recurring agent runs | Wake the harness without a user prompt. |
| s23 Audit & Sandbox | Command policy and tamper-evident logs | Make autonomous work governable. |
| s24 Comprehensive | End-to-end mini harness | See how all layers fit together. |
The tutorial is a ladder:
agent loop
+ tool dispatch
+ permissions
+ desktop process boundary
+ sidecar/session runtime
+ memory/context compression
+ extension protocols
+ persistence/scheduling/audit
= WorkBuddy-style harness
Each chapter should answer three questions:
- What problem appears when the previous chapter gets real users?
- What minimal mechanism fixes it?
- How would a production desktop agent harden that mechanism?
社区求职指南（AgentGuide）把 Agent 工程面试归纳为 8 个考点。本教程每个考点都有对应的可运行章节——学完一章，就多一个能在面试里从"讲概念"升级到"讲我怎么实现的"的话题：
| 考点 | 对应章节 | 面试时你能讲的实现细节 |
|---|---|---|
| Agent Loop | s01, s02 | tool_use/tool_result 反馈环、stop_reason 驱动的退出条件、工具分发表 |
| Context & Cost | s03, s13, s14 | 工具 schema 延迟加载、大输出外部化（swap file + 指针）、token 压力检测与结构化压缩 |
| Tool Design | s02, s03, s17 | 路径守卫、schema 设计、ToolSearch 按需发现、MCP 连接器命名空间 |
| Multi-Model Routing | s08 | lite/default/craft 三档路由、成本追踪、Agent 到模型槽位的映射 |
| Memory | s10, s11, s12 | 工作区日志→主题蒸馏、用户偏好去重、三层记忆的所有权边界 |
| Eval & Governance | s23, tests/ | 哈希链审计（含截断锚点）、命令安全分级、权限门单元测试与绕过边界用例 |
| Harness | s05, s06, s07, s24 | UI/sidecar/session 三层进程隔离、JSON-RPC 控制面、端到端 mini harness |
| Reliability | s09, mini_workbuddy | JSONL append-only 转录与崩溃恢复、fail-closed 错误边界、可靠工具 call id |
配套阅读顺序见 Further Reading Map。