Technical book · sneak peek
Agents with Foresight
World Models for Reliable Action in Digital Enterprises
An agent can say nothing wrong and still act on an account that no longer exists.
Agents now write to business records through tools and other agents. The enterprise keeps moving underneath them. Payments post late. Policies change in the middle of a conversation.
This book is about the distance between what an agent holds and what the enterprise has become. It follows one synthetic incident from the first read to the post-incident review. Each chapter adds one instrument and runs it as standard-library Python.
Why it matters
One question after every incident
After an incident the logs show every word the agent said. They show nothing about what it believed. A stale read and a receipt taken for completion leave the same trail.
This book exists for one question. What did the system believe when it acted?
Where world models fit
Prediction has to earn its place
The claim is narrow on purpose. Most failures of this kind are cured by a fresh read and a conditional write. Prediction earns its place only when comparing futures improves a decision. The gain has to pay for model error, latency and maintenance.
World model means two things right now. In research it is a trained network that predicts the next state of the world.
In this book it is any representation that supports a prediction about the world a decision depends on. That includes how the world responds to the agent's own action. The book teaches you to tell which case you are in.
What you will learn
Four things you can do after the last chapter
- Decide whether a decision needs a predictive model or only a fresh read and a version check
- Wrap each observation in an envelope with its source, version and four clocks
- Put a version-checked action boundary between the agent and the record
- Reconstruct an incident and answer what the system believed when it acted
Who it is for
Builders who own the write path
For senior software engineers, AI engineers and solution architects whose agents act on systems of record.
Excerpt
From the Preface
At minute nine, an agent called Marlow submits an installment plan for account 88-1204. The service answers
accepted, so Marlow tells the member the plan is set up and closes the task. Nothing Marlow said was wrong, yet the account it described had stopped existing seven minutes earlier.The plan rested on a balance of 600 units that Marlow read at minute two. A payment of 100 had already been accepted by the processor but not yet posted to the ledger. The installment policy changed at minute six and nobody told Marlow.
At minute forty-five, Ines pulls the logs. She can see every word Marlow said and nothing about what it believed. A stale read, a policy race and a receipt taken for completion all leave the same trail.
This book exists for one question. What did the system believe when it acted?
A short excerpt from the manuscript in revision. Wording may change before publication.
The author
Varma (FNU Ratnavarma Kundady)
Varma has 21 years in enterprise technology, more than ten of them in the utility sector. His research covers agent memory and governed execution. It also covers how to evaluate agents that act on business records.
Read alongside
The World Outside the Model Engineering AI Agents for State, Authority and Reliable Action A companion volume. It covers durable execution and runtime authority.