---
theme: seriph
title: "Lesson 01 — How Do We Replace Agent Intuition with Evidence?"
info: "English video course for AI Agents in Depth"
author: Bojie Li
transition: slide-left
mdc: true
lineNumbers: false
monaco: false
aspectRatio: 16/9
canvasWidth: 980
layout: cover
class: cover
---
Build · Introduction · Orientation
# How Do We Replace Agent Intuition with Evidence?
A practice-first map of AI Agents in Depth
Lesson 01 of 42 · 16 minutes · Introduction; Book Structure; How to Read This Book
---
layout: center
class: text-center
---
The central question
Why do impressive Agent demos so often fail to become reliable products?
---
# Why this problem matters
Demo success
One lucky trajectory proves possibility—not reliability.
Engineering judgment
Every design choice needs a mechanism and a trade-off.
Scientific progress
Without evaluation, change is indistinguishable from luck.
---
# Three ideas to keep in view
Build
Context, knowledge, tools, and code generation
Improve
Evaluation, post-training, and continual evolution
Expand
Voice, Computer Use, robotics, and collaboration
---
# The book's visual model
The four-part structure of the book
---
# Demo-driven vs. Principle-driven
Demo-driven
- Start with a framework
- Celebrate one successful run
- Change prompts by intuition
Principle-driven
- Start with a failure mode
- Run a controlled comparison
- Turn evidence into a design rule
The course follows the right-hand loop.
---
# The course's experimental loop
~~~python
question = define_failure_mode()
hypothesis = predict_mechanism(question)
evidence = run_controlled_experiment(hypothesis)
rule = interpret(evidence, limitations=True)
evaluate(rule)
~~~
---
# Test the claim
Course tour1 min
Inspect one companion experiment before running it
Observe: Entry point, modes, providers, outputs, and reproducibility controls
Demo budget: 1 minute · one contiguous terminal block
---
class: course-terminal
---
Live demo
# Switching to the terminal
~~~bash
$ uv run python chapter1/context/main.py --help
~~~
Run the command(s), narrate decisions, and point to the observation—not just the output.
---
# What the evidence supports
Finding 1
The book is organized around recurring engineering questions, not products.
Finding 2
Experiments expose mechanisms through controls, ablations, and receipts.
Finding 3
The author's interpretation—not terminal output alone—is the course's value.
---
layout: center
---
Where the claim stops
# Boundary condition
A short lesson cannot reproduce every long-running campaign. It can make the protocol and evidence traceable.
---
layout: center
---
Engineering takeaway
# Design rule
Never present an Agent result without first stating what would count as success or failure.
---
# Continue the experiment
---
layout: center
class: text-center
---
Pause and apply
# Your turn
Which Agent claim have you accepted after seeing only one successful run?
---
layout: center
class: text-center
---
Introduction complete · Next · Lesson 02
Define an Agent by the interfaces that connect it to the world.
→