01
Model
Proposes decisions. It is a component, not the whole system.
A code-first book about useful autonomy
Agentic Engineering
from First Principles
Building, Controlling, and Scaling LLM Agents in Practice
Build the smallest possible agent. Run it. Break it deliberately. Then engineer the failure.
30chapters
59experiments
1tiny agent kernel
CONCEPT→CODE→RUN→BREAK→ENGINEER→PATTERN→SCALE
The central model
Strip away the terminology.
An agent is a software system in which a model participates in decisions that can affect an environment. Most of the field can be understood through five components.
Agent = Model + Context + Actions + State + Control
02
Context
The information selected and made visible for the current decision.
03
Actions
Capabilities whose consequences can change an external environment.
04
State
What persists: working state, memory, artifacts, checkpoints and experience.
05
Control
What may happen, how execution proceeds, what resources are available, and when it stops.
THE CONTROL LAWAutonomy ↑ ⇒ Control Requirements ↑
The objective is not maximum autonomy. It is useful autonomy under appropriate control.
The journey
30 chapters. Six parts.
Start with one model call and progressively add abilities, reasoning, collaboration, reliability, and advanced autonomy.
Executable learning
59 experiments. One question at a time.
Every experiment isolates a mechanism, creates a failure, and turns the failure into an engineering lesson.
No magic
The smallest useful agent fits on one screen.
Start with a visible control loop, then compare all 59 experiments with runnable LangChain and LangGraph implementations.
- Provider-neutral model interface
- Explicit tool execution
- Bounded steps and stopping
- Deterministic examples before empirical benchmarks
- Primitive → Engineered → Framework
agent.py
class Agent:
def run(self, goal):
messages = [goal]
for step in range(10):
decision = self.model(messages)
if decision.done:
return decision.answer
result = self.tools[
decision.tool
](**decision.arguments)
messages.append(result)
raise RuntimeError(
"Step limit reached"
)The complete book
All 30 chapters.
Yours to explore.
Download the complete 381-page PDF, including the cover, first-principles and framework examples, and the full journey from a model call to controlled autonomy.
- Enter your email address.
- Confirm it using the link in your inbox.
- Download the complete book.
You can still read Chapter 1 immediately.
Principles revealed by the experiments
More is not automatically better.
More Context ≠ Better Context
More Reasoning ≠ Better Answers
More Memory ≠ Better Memory
More Agents ≠ Better Intelligence
More Autonomy ≠ Better Systems
Adaptation ≠ Improvement
Outcome Quality ≠ Trajectory Quality
Capability ≠ Permission
Consensus ≠ Truth
About the author
Christiaan van der Walt, PhD
Christiaan van der Walt is an engineer, machine learning researcher, and technology leader with nearly two decades of experience developing AI systems from first principles and bringing them into production. He currently leads Advanced Analytics within Momentum Group’s Group Technology structure. His work spans generative AI, recommendation systems, predictive analytics, and patented digital fitness assessment technology.
His earlier career includes several years as a machine learning researcher at the Council for Scientific and Industrial Research (CSIR), where he developed algorithms and systems across speech processing, biometrics, security, and environmental modelling. He also completed a research internship at IBM’s T.J. Watson Research Center in New York, working on automatic speech recognition and speech-driven information retrieval.
Christiaan holds a PhD specialising in machine learning from North-West University and engineering degrees from the University of Pretoria. His research includes contributions to density estimation and pattern recognition, with publications at conferences including AAAI and ICASSP. His experience combines mathematical foundations, hands-on software development, research, and the practical demands of deploying AI in real organisations.
Start at the beginning
Build the smallest possible agent.
Then break it.
59 executable experiments. One tiny agent kernel. No magic.