ai-agent-book/slides/lesson-21.md

theme, title, info, author, transition, mdc, lineNumbers, monaco, aspectRatio, canvasWidth, layout, class

theme	title	info	author	transition	mdc	lineNumbers	monaco	aspectRatio	canvasWidth	layout	class
seriph	Lesson 21 — How Can Code Let an Agent Create New Capabilities?	English video course for AI Agents in Depth	Bojie Li	slide-left	true	false	false	16/9	980	cover	cover

Build · Chapter 5 · Coding Agents

How Can Code Let an Agent Create New Capabilities?

Adapters, generative UI, hot repair, and Agent bootstrapping

Lesson 21 of 42 · 19 minutes · Code as a System Adapter; Generative UI; Agent Bootstrapping

---

Why this problem matters

Adapter

Translate unstable external formats into stable internal ones.

Interface

Generate a UI that matches the current intent.

Bootstrap

Create a specialized Agent from a validated reference.

---

Three ideas to keep in view

Hot repair

Failure sample → generated parser → tests → registration

Artifact pattern

Pass paths and queries instead of moving large data through tokens

Validated generation

Compile, test, scan, and exercise generated Agents

---

The book's visual model

Pipeline of an Agent that creates Agents

---

One-off generation vs. Capability creation

One-off generation

• Produces an artifact
• Capability disappears after task
• No reuse gate

Capability creation

• Packages implementation
• Validates and versions it
• Reuses on later tasks

Bootstrapping begins when output becomes part of the next Agent.

---

Capability promotion needs gates

candidate = agent.generate_tool(failure_sample)
compile(candidate)
run_security_scan(candidate)
run_regression_tests(candidate)
registry.promote(candidate)

---

Test the claim

5-72 min

Repair an unknown log format

Observe: Failure detection, generated parser, tests, and later reuse

5-92 min

Generate a dynamic form

Observe: Intent gaps converted into fields and cascading constraints

5-122 min

Inspect Agent-creation validation gates

Observe: Scratch versus template generation, compile, tests, and protocol audit

Demo budget: 6 minutes · one contiguous terminal block

---

class: course-terminal

Live demo

Switching to the terminal

$ uv run python chapter5/adaptive-log-parser/demo.py --offline

$ uv run python chapter5/dynamic-form/demo.py --offline

$ uv run python chapter5/agent-creator/demo.py --no-live --output chapter5/agent-creator/runs/course-smoke

Run the command(s), narrate decisions, and point to the observation—not just the output.

---

What the evidence supports

Finding 1

Generated adapters let a system follow changing data formats.

Finding 2

Generative UI moves structured clarification out of slow chat turns.

Finding 3

Reference-based Agent creation preserves a proven loop while specializing tools.

---

Boundary → design rule

Self-modification is unsafe when the Agent can alter the validator that approves its own changes.

Promote generated code into capability only after independent security, functional, and reuse checks.

---

Continue the experiment

Experiment 5-8: log diagnosis chapter5/log-diagnosis/ Experiment 5-10: ERP SQL Agent chapter5/erp-agent/ Experiment 5-11: conversational UI chapter5/conversational-ui/ Dynamic form architecture book-en/images/fig5-8.svg SQL artifact pattern book-en/images/fig5-9.svg

---

layout: center class: text-center

Pause and apply

Your turn

Which generated artifact is safe to reuse, and who decides that it has become a capability?

---

layout: center class: text-center

Build complete · Next · Lesson 22

Measure whether any of these architectural changes actually improve the Agent.

→