Part 2 — For the Developer

Ch. 18 — Incident Response for Agentic Systems

What to do when an agent does something wrong — detection, containment, root cause, and regulatory notification.

When a traditional system fails, you find a stack trace. When an agent does something wrong, you may not find out until the downstream harm has already happened.

Standard incident response was designed for deterministic failure modes. Agents fail differently. Before you can build a response playbook, you need to understand what you are responding to.

Scope violation — the agent accessed data or took an action outside its defined boundaries. The cause may be a prompt injection attack, a bug in tool routing, or access controls that were never properly configured. Whatever the cause, the result is an agent doing something it was never authorized to do.

Platform Agentic

Compliance, governance, and accountability for teams building agentic AI systems.

Access the book — sign in with Google·LinkedIn