Computational Judgment

AI agents are already producing moments that feel almost magical. They write code, investigate failures, analyze evidence, use tools, and find useful ways forward through situations no one programmed step by step.

Teams are now trying to turn those impressive moments into reusable AI capability. They add better instructions, examples, skills, tools, data, safeguards, and human support so that useful performance can recur across new situations.

But as these systems grow, a fundamental question becomes harder to answer:

What kind of AI capability are we actually building?

Enterprises are building AI capability faster than they can understand what they are creating.

Computational Judgment offers a new way to see the problem. Its central claim is simple:

A model is trained. An agentic system develops.

The book begins with the familiar experience of teaching an AI agent-correcting it, showing it examples, giving it better information and tools, and watching its work improve. Some of that support can be made reusable. Yet more guidance does not always produce more capability, and the same model can perform brilliantly in one situation and poorly in another.

The book introduces computational judgment as a way to understand what happens when an agentic system works through situations whose path cannot be fully specified in advance.

From this perspective, AI capability is larger than the model, prompt, skill, or workflow. It can be deliberately built, improved through evidence, demonstrated across real situations, and responsibly governed.

Written for AI engineers, architects, product leaders, researchers, and anyone building serious enterprise AI systems, Computational Judgment provides a foundation for moving from impressive AI performance to capabilities that can be understood, developed, and trusted.

The next frontier is not better prompting. It is learning how to develop computational judgment.