Guardrails for AI Agents

Most agent tutorials show you how to make an agent act. This one shows you how to stop it.

A model that can issue refunds, send messages or update records will eventually try to do the wrong one. The usual defence is a sentence in a prompt, and a sentence in a prompt is a request, not a rule. This book takes a different approach, built on one idea:
The model proposes. The guardrail decides. The tool executes.

You build nine small, working LangGraph agents, each adding a guardrail at a new point:

• Input — clean sensitive data before the model ever sees it
• Action — check a tool call before the tool runs
• Risk routing — send low, medium and high risk down different paths
• Human approval — pause the agent mid run and let a person decide
• Memory — stop the wrong things being remembered forever
• Retrieval — filter documents by permission, and catch instructions hidden inside them
• Plan — approve a whole sequence before the first step runs
• Output — read the answer before the customer does
• Blast radius — catch the hundred small actions that add up to one large problem

Every guardrail is plain Python: predictable, testable in milliseconds without an API key, and impossible to talk out of doing its job. Every chapter opens with a real failure, shows the code that prevents it, tests it, and then tells you honestly what it still does not catch.

Who this is for

Engineers building or reviewing agents that do something real. You need to read Python. You do not need LangGraph or a security background. The policy files and tests are plain enough to show to a manager, a risk officer or an auditor, and several chapters suggest doing exactly that.

What you get

• 12 complete chapters
• A companion repository with all nine programs and a full test suite
• A one page reference card listing every guardrail in a single line

Still growing

Part III, Production Guardrails, is being added one chapter at a time: trust boundaries, authorization and budgets, idempotency, classifier based checks, grounding, and a shared policy gateway. Everyone who owns the book gets every update at no extra cost.

The examples run on Claude, and every guardrail is plain Python that works with any model.

Companion repository: github.com/sameershukla/guarded-agents-in-action