StackPilot

Infrastructure changes rarely fail because a team cannot write the code. They fail because intent is incomplete, evidence is scattered, consequences are hidden, and authority is granted without binding it to the exact change that was reviewed.

StackPilot presents a practical architecture for using AI in infrastructure operations without turning a language model into an unchecked production operator.

The book begins with a deceptively simple request: give a struggling service enough capacity to meet a deadline. What appears to be a small configuration adjustment quickly becomes a decision involving cost, security, dependencies, ownership, timing, recovery, and live operational state.

From that starting point, the book develops a graph-grounded approach to governed infrastructure change.

You will learn how to:

• translate ambiguous requests into typed change-intent records;
• retrieve evidence with provenance, authority, freshness, and tenant scope;
• model dependencies and blast radius with a typed change graph;
• convert infrastructure plans into consequence-oriented review packets;
• apply deterministic policy without confusing compliance with judgment;
• bind human approval to an exact, expiring execution subject;
• execute changes through isolated, short-lived cells with narrow workload identities;
• observe the live result, detect partial success or drift, and reconcile outcomes;
• evaluate AI assistance without rewarding fluency over correctness; and
• introduce production authority gradually through measurable milestones.

The central principle is simple:

AI proposes. Controls authorize. Isolated cells execute. Observation closes the loop.

StackPilot is not a Terraform primer, a catalog of cloud services, or a promise of fully autonomous operations. It is a design guide for platform engineers, SREs, cloud and security architects, engineering leaders, and applied-AI practitioners who want AI assistance that remains inspectable, bounded, and accountable.

The book combines operational stories, architecture diagrams, formal models, worked traces, review questions, exercises, implementation guidance, and a field guide for designing the first governed change.

By the end, you will have a practical blueprint for building one complete, reviewable path from request to evidence, plan, authorization, execution, and observed outcome—without granting an AI system more authority than it has earned.