Agent Patterns

AI agents extend LLM capabilities with tools, memory, and multi-step reasoning. The industry frames this as capability enhancement.

The security framing is different: every tool call is an authority decision made by a system that can't verify intent. Every action loop is an opportunity for injection to compound. Every multi-agent interaction is cross-principal trust confusion.

These patterns document how agents are built, where the authority model breaks down, and what you need to do to build them securely.