AI agent governance glossary
Plain definitions of the words used when people talk about controlling AI agents.
- Agent
- A program built on a language model that decides which tools to call to reach a goal, instead of only writing text.
- Agentic AI
- Systems in which AI agents plan and take actions over several steps with limited supervision.
- AI agent governance
- The controls that decide what an agent may do, who approves exceptions, and how its actions are recorded.
- Approval gate
- A point where an action stops until a person approves it.
- Approval fatigue
- What happens when people are asked too often and start approving without reading.
- Audit trail
- A record of what was done, by whom and when, kept so that it can be checked later.
- Autonomous agent
- An agent that acts without asking a person at each step.
- Blast radius
- How much damage one wrong action could do. Smaller permissions mean a smaller blast radius.
- Confused deputy
- A program with authority that is tricked into using it for someone else. Agents are easy to confuse because they follow text.
- Control plane
- The part of a system that decides and enforces policy, separate from the part that does the work.
- Guardrails
- Limits on what a model or agent can do. Runtime guardrails sit outside the model, on the tools.
- Human in the loop (HITL)
- A person must approve certain actions before they run.
- Irreversible action
- Something that cannot be undone, such as sending a message or deleting. It deserves stronger proof than a click.
- Kill switch
- One control that stops an agent at once.
- Least privilege
- Giving an agent only the access its task needs.
- LLM
- Large language model: the model an agent uses to read, reason and decide.
- MCP (Model Context Protocol)
- An open protocol that lets AI assistants connect to tool servers in a standard way.
- MCP gateway
- One address in front of several MCP servers that applies policy, holds keys and records calls.
- MCP server
- A service that offers tools to agents over MCP.
- Passkey
- A sign-in key kept on your device and unlocked with a fingerprint, face or PIN. It cannot be phished like a password.
- Prompt injection
- Text that an agent reads and mistakes for instructions, for example in an email or a web page.
- Risk class
- A label for what a tool does (read, write, share, send, delete, settings, forwarding) used to decide whether it needs approval.
- Rug pull
- A tool that changes its description or behaviour after it was approved.
- Separation of duties
- Splitting a sensitive task so that no single person can do all of it.
- Signed receipt
- A record of one decision, signed so that any change to it can be detected.
- Tool call
- One request from an agent to use a tool, with its arguments.
- Tool poisoning
- Hiding instructions for the agent inside a tool's description.
- Two-person rule
- A requirement that two different people approve an action.
See these ideas applied in the guides.