Defensive Tool Design: The Key to Building Better AI Agents

AI agents become useful when they can do more than generate text. They search internal systems, retrieve documents, update records, write code, send messages, create bookings, and trigger business processes. Each tool makes the agent more capable, but it also creates a new way for a plausible model mistake to become a real-world incident. Defensive Tool Design is a practical guide to building that capability boundary well. Rather than treating prompts as security controls, the book shows how to put safety and reliability into the tool itself: its purpose, schema, permissions, side effects, failure modes, approval flow, retry behaviour, and audit trail. It explains why identity should come from trusted runtime context, why destructive actions should be separated into proposal and commit stages, why read-only tools can still be dangerous, and why a constraint that is not enforced or evaluated is not a guardrail. >For engineers moving from agent demos to production systems, Defensive Tool Design provides a disciplined way to decide what an agent should be allowed to do-and how to make sure it cannot quietly do more.