Files
liqiang b119135836
Build latest book artifacts / build (push) Canceled after 0s
dependency resolution / resolve (3.11) (push) Canceled after 0s
dependency resolution / resolve (3.13) (push) Canceled after 0s
deploy-pages / build (push) Canceled after 0s
deploy-pages / deploy (push) Canceled after 0s
i18n consistency check / check (push) Canceled after 0s
provider adoption tests / test (chapter2/context-compression) (push) Canceled after 0s
provider adoption tests / test (chapter2/prompt-injection) (push) Canceled after 0s
provider adoption tests / test (chapter2/system-hint) (push) Canceled after 0s
provider adoption tests / test (chapter3/log-sanitization) (push) Canceled after 0s
web-search-agent tests / test (push) Canceled after 0s
web-search-agent tests / agentbook (push) Canceled after 0s
ai-agent-book 精选快照(<2MB 代码与文档,来自 github.com/bojieli/ai-agent-book)
2026-08-20 13:12:50 +00:00

139 lines
11 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# AI Agents in Depth — English Video Course
Approved Option B curriculum: 42 problem-oriented lessons following the English book order. Each lesson is 1520 minutes, budgeting approximately one minute per Slidev slide plus one to three minutes per live experiment.
## Learning arc
| Movement | Chapters | Viewer progression |
| --- | --- | --- |
| Build an Agent | IntroductionChapter 5 | Context → memory → tools → executable capabilities |
| Improve it scientifically | Chapters 68 | Evaluation → post-training → continual evolution |
| Expand it | Chapters 910 | Voice → Computer Use → robotics → multi-Agent collaboration |
## Approved chapter allocation
| Book section | Lessons | Count |
| --- | ---: | ---: |
| Introduction | 1 | 1 |
| Chapter 1 | 24 | 3 |
| Chapter 2 | 59 | 5 |
| Chapter 3 | 1013 | 4 |
| Chapter 4 | 1417 | 4 |
| Chapter 5 | 1821 | 4 |
| Chapter 6 | 2225 | 4 |
| Chapter 7 | 2631 | 6 |
| Chapter 8 | 3234 | 3 |
| Chapter 9 | 3538 | 4 |
| Chapter 10 | 3942 | 4 |
| **Total** | **142** | **42** |
Chapter 7 intentionally receives six lessons because post-training and reinforcement learning are the largest conceptual jump for viewers without prior ML-training knowledge. Chapter 9 receives four lessons so Computer Use and robotics have separate mechanisms, experiments, and safety boundaries.
## Lesson-by-lesson outline
### Introduction
| Lesson | Problem-oriented title | Main learning outcome | Live anchor(s) | Slides | Demo | Target |
| ---: | --- | --- | --- | ---: | ---: | ---: |
| [01](lesson-01.md) | How Do We Replace Agent Intuition with Evidence? | A practice-first map of AI Agents in Depth | Course tour | 15 | 1 min | 16 min |
### Chapter 1
| Lesson | Problem-oriented title | Main learning outcome | Live anchor(s) | Slides | Demo | Target |
| ---: | --- | --- | --- | ---: | ---: | ---: |
| [02](lesson-02.md) | What Makes an AI System an Agent? | Reasoning engine + working context + action interfaces | 1-2 | 15 | 2 min | 17 min |
| [03](lesson-03.md) | Why Does an Agent Need Its Entire Trajectory? | ReAct, context components, and systematic ablation | 1-1A, 1-1B | 15 | 4 min | 19 min |
| [04](lesson-04.md) | Why Doesn't a Stronger Model Make a Reliable Agent? | Harness engineering, orchestration, and guardrails | 1-3 | 15 | 2 min | 17 min |
### Chapter 2
| Lesson | Problem-oriented title | Main learning outcome | Live anchor(s) | Slides | Demo | Target |
| ---: | --- | --- | --- | ---: | ---: | ---: |
| [05](lesson-05.md) | What Does the Model Actually See? | Messages, tool calls, and the Agent core loop | 2-1 | 15 | 3 min | 18 min |
| [06](lesson-06.md) | Why Can One Timestamp Make an Agent Slow? | Chat templates, attention, KV Cache, and stable prefixes | 2-3, 2-2 | 15 | 4 min | 19 min |
| [07](lesson-07.md) | Why Do Better Prompts Need Structure, Not More Rules? | Process-oriented instructions, tool definitions, and injection boundaries | 2-4, 2-5 | 13 | 5 min | 18 min |
| [08](lesson-08.md) | How Can an Agent Know What It Needs to Learn? | Skills, progressive disclosure, and on-demand capability | 2-6 | 15 | 3 min | 18 min |
| [09](lesson-09.md) | How Can an Agent Stay Oriented in a Long Task? | Status bars, physical time, context rot, and compression | 2-8, 2-9 | 15 | 4 min | 19 min |
### Chapter 3
| Lesson | Problem-oriented title | Main learning outcome | Live anchor(s) | Slides | Demo | Target |
| ---: | --- | --- | --- | ---: | ---: | ---: |
| [10](lesson-10.md) | What Should an Agent Remember About a User? | Memory levels, representations, evaluation, and privacy | 3-1/3-2, 3-3 | 14 | 5 min | 19 min |
| [11](lesson-11.md) | Why Does Semantic Search Miss Exact Answers? | Chunking, dense retrieval, sparse retrieval, and evaluation | 3-4, 3-5 | 15 | 4 min | 19 min |
| [12](lesson-12.md) | Why Is One Retrieval Index Never Enough? | Hybrid search, reranking, multimodality, and structured knowledge | 3-6, 3-8 | 13 | 5 min | 18 min |
| [13](lesson-13.md) | When Should the Agent Decide What to Retrieve? | Agentic RAG, contextual retrieval, and two-tier memory | 3-9, 3-11, 3-12 | 13 | 6 min | 19 min |
### Chapter 4
| Lesson | Problem-oriented title | Main learning outcome | Live anchor(s) | Slides | Demo | Target |
| ---: | --- | --- | --- | ---: | ---: | ---: |
| [14](lesson-14.md) | What Makes a Tool Easy for a Model to Use? | Capability boundaries, granularity, descriptions, and MCP | 4-1 | 15 | 3 min | 18 min |
| [15](lesson-15.md) | How Do You Let an Agent Act Without Letting It Cause Damage? | Execution tools, independent checks, and fail-closed design | 4-3A, 4-3B | 15 | 3 min | 18 min |
| [16](lesson-16.md) | When Should an Agent Ask for Help or Delegate? | Sub-agents, Human-in-the-Loop, and communication tools | 4-4A, 4-4B | 15 | 3 min | 18 min |
| [17](lesson-17.md) | How Can a Synchronous Model Live in an Asynchronous World? | Events, interruption, parallelism, and proactive tool discovery | 4-5, 4-6, 4-7 | 13 | 6 min | 19 min |
### Chapter 5
| Lesson | Problem-oriented title | Main learning outcome | Live anchor(s) | Slides | Demo | Target |
| ---: | --- | --- | --- | ---: | ---: | ---: |
| [18](lesson-18.md) | Why Is Code Generation Not Enough to Build a Coding Agent? | Files, execution, harness recovery, and bounded verification | Coding workflow, Harness tests | 15 | 3 min | 18 min |
| [19](lesson-19.md) | When Should an Agent Think in Code Instead of Words? | Math, logic, and deterministic business constraints | 5-1, 5-2, 5-3 | 13 | 6 min | 19 min |
| [20](lesson-20.md) | How Can an Agent Create Media It Can Actually Verify? | Slidev, rendering, multimodal review, and video editing | 5-4, 5-6 | 13 | 5 min | 18 min |
| [21](lesson-21.md) | How Can Code Let an Agent Create New Capabilities? | Adapters, generative UI, hot repair, and Agent bootstrapping | 5-7, 5-9, 5-12 | 13 | 6 min | 19 min |
### Chapter 6
| Lesson | Problem-oriented title | Main learning outcome | Live anchor(s) | Slides | Demo | Target |
| ---: | --- | --- | --- | ---: | ---: | ---: |
| [22](lesson-22.md) | How Do You Test an Agent Instead of Its Final Answer? | Environments, state, datasets, and executable verification | Evaluation control | 16 | 2 min | 18 min |
| [23](lesson-23.md) | How Do You Judge Quality Without Hiding Failure? | Rubrics, vetoes, LLM judges, pairwise comparison, and Elo | 6-3, 6-6 | 15 | 4 min | 19 min |
| [24](lesson-24.md) | Which Agent Should You Ship? | Model behavior, latency, cost, and evaluation-driven selection | 6-8, 6-7 | 15 | 4 min | 19 min |
| [25](lesson-25.md) | Did the Agent Improve—or Did the Numbers Move? | Significance, observability, ablations, and production evaluation | 6-3 evidence | 15 | 2 min | 17 min |
### Chapter 7
| Lesson | Problem-oriented title | Main learning outcome | Live anchor(s) | Slides | Demo | Target |
| ---: | --- | --- | --- | ---: | ---: | ---: |
| [26](lesson-26.md) | Why Does Model Training Happen in Three Stages? | Pre-training, SFT, RL, and the agent-environment loop | 7-1 | 15 | 2 min | 17 min |
| [27](lesson-27.md) | When Should You Teach with Examples—and When with Rewards? | SFT, loss masking, distribution shift, and the form-first rule | 7-4 evidence, 7-5 evidence | 15 | 4 min | 19 min |
| [28](lesson-28.md) | How Do Preferences Become a Trainable Signal? | RLHF, reward models, KL constraints, PPO, GRPO, and DPO | RL evaluation check | 15 | 2 min | 17 min |
| [29](lesson-29.md) | Why Do Data and Environments Matter More Than the Algorithm? | Practice grounds, task distributions, synthetic data, and fidelity | 7-9 data | 15 | 2 min | 17 min |
| [30](lesson-30.md) | How Do You Reward a Long Agent Trajectory? | Credit assignment, reward density, process signals, and path penalties | 7-14 gates | 15 | 2 min | 17 min |
| [31](lesson-31.md) | How Can a Model Learn to Use Tools with Fewer Samples? | Tool-call RL, sandbox feedback, distillation, and practical boundaries | 7-9 preflight | 15 | 2 min | 17 min |
### Chapter 8
| Lesson | Problem-oriented title | Main learning outcome | Live anchor(s) | Slides | Demo | Target |
| ---: | --- | --- | --- | ---: | ---: | ---: |
| [32](lesson-32.md) | How Do Failed Trajectories Become Learning Signals? | Outcome verification, process rules, Rubrics, and cross-trajectory experience | 8-1, 8-2 | 15 | 4 min | 19 min |
| [33](lesson-33.md) | Where Should an Agent Store What It Learns? | Knowledge, instructions, programs, parameters, and meta-updates | 8-4, Tool evolution | 15 | 4 min | 19 min |
| [34](lesson-34.md) | How Can a Self-Modifying Agent Change Without Drifting? | Candidate gates, transfer, retention, rollback, and sleep learning | 8-5, 8-6 | 15 | 4 min | 19 min |
### Chapter 9
| Lesson | Problem-oriented title | Main learning outcome | Live anchor(s) | Slides | Demo | Target |
| ---: | --- | --- | --- | ---: | ---: | ---: |
| [35](lesson-35.md) | Why Does a Voice Agent Feel Slow? | Cascaded pipelines, latency waterfalls, streaming, and turn detection | 9-1, 9-2 | 15 | 3 min | 18 min |
| [36](lesson-36.md) | When Should Voice Stop Taking Turns? | Omni, full-duplex interaction, fast-slow thinking, and controllable speech | 9-3, 9-4 | 15 | 4 min | 19 min |
| [37](lesson-37.md) | How Does an Agent Act Through Pixels? | GUI action spaces, visual grounding, and bounded interaction | 9-6 preflight, 9-6 retained status | 15 | 3 min | 18 min |
| [38](lesson-38.md) | How Does an Agent Turn Plans into Physical Actions? | Planning-control separation, VLA control, safety gates, and Sim2Real | 9-9 dry configuration, Robot safety gates | 15 | 3 min | 18 min |
### Chapter 10
| Lesson | Problem-oriented title | Main learning outcome | Live anchor(s) | Slides | Demo | Target |
| ---: | --- | --- | --- | ---: | ---: | ---: |
| [39](lesson-39.md) | When Should Agents Share the Same Context? | Shared trajectories, isolated contexts, role switching, and handoffs | 10-1, 10-2 | 15 | 2 min | 17 min |
| [40](lesson-40.md) | Who Should Coordinate Independent Agents? | Peer review, managers, decentralized handoffs, files, and control planes | 10-3 | 15 | 2 min | 17 min |
| [41](lesson-41.md) | When Is Multi-Agent Actually Better Than One Agent? | Information gain, parallelism, verification, budgets, and cost | 10-6 | 15 | 3 min | 18 min |
| [42](lesson-42.md) | How Do Agent Teams Fail—and What Should We Build Next? | Conflicts, error cascades, Agent societies, and the course synthesis | 10-8 offline diagnostic | 16 | 2 min | 18 min |
## Recording contract
- Speak in your own voice and add interpretation; the decks are visual prompts, not narration scripts.
- Run the listed commands in one contiguous terminal block after the explicit handoff slide.
- Treat preflights, validators, smoke checks, and dry configurations as scoped evidence—not completed long campaigns.
- Use the linked companion projects for experiments that are not demonstrated live.
- Demo-heavy lessons combine or remove conceptual slides so slide time plus terminal time stays within 20 minutes.