MCP primitives are the five core building blocks of the Model Context Protocol that define how AI agents connect to external tools and data sources.

A subagent is a separate AI agent instance with its own context window that handles a scoped task independently, acting as both a specialization tool and a context pressure valve.

AGENTS.md is a repository-level markdown file that gives AI coding agents the project context they need to write correct code from the very first session.

Progressive disclosure is a three-level loading strategy that feeds AI agents only the context they need at the moment, keeping sessions fast and token costs low.

Agent skills are reusable markdown files that teach AI coding tools how to perform repeated tasks consistently, replacing ad-hoc prompts with version-controlled procedures.

Blast radius is the total scope of damage a compromised or malfunctioning AI agent can cause before anything stops it, and limiting it is the most practical security strategy you have.

Tool poisoning is a stealth attack that embeds malicious instructions inside MCP tool descriptions, corrupting your AI agent before a single user command is typed.

Slopsquatting is a supply chain attack where adversaries register package names that AI coding tools hallucinate, turning a simple install command into a security breach.

The lethal trifecta describes the dangerous combination of private data access, untrusted content exposure, and outbound action capability that makes AI agents exploitable.