Reference

The vocabulary of agent authority.

Short, operational definitions for the terms that matter when AI agents get real tools: runtime authorization, MCP security, prompt injection, human review, and policy-as-code. Each entry explains where the concept lands in production.

Use this glossary

The terms below are organized by the part of the agent system they touch. If you came here from a policy file, a docs page, or a security write-up, pick the matching category. If you are doing a security review, read Authorization and Agent Security back-to-back. Cross-links at the bottom of each entry connect to product pages, compliance briefs, and integration guides.

Authorization

Concepts that govern what an agent is allowed to do at the moment it tries to do it.

Runtime Authorization

Authorization decisions made at action time against current context, not at deploy time.

Agent Authorization

Deciding what an agent identity is allowed to do, separate from the human who started it.

Tool-Call Authorization

Intercepting each tool invocation and deciding allow, block, or require-approval before execution.

The Authorization Gap

The space between being signed in and what an agent should be allowed to do with a tool.

MCP

Model Context Protocol security and gateway patterns.

MCP Gateway

A proxy that sits in front of MCP servers and enforces policy on each governed tool call.

MCP Tool Poisoning

Malicious or compromised tool definitions that smuggle instructions into the agent.

Agent Security

Threats and attack patterns that target LLM-driven agents.

Prompt Injection

Manipulating an LLM by smuggling instructions into its input. OWASP LLM01.

Indirect Prompt Injection

Hidden instructions embedded in content the agent reads, like a webpage or a document.

OWASP LLM06: Excessive Agency

Top-ten risk for LLM apps: an agent has more capability than the situation justifies.

Excessive Agency

When an agent's capability set exceeds its actual authority. The root cause behind most agent incidents.

Patterns

Human oversight and approval patterns adapted from finance and ops for AI agents.

Human Veto Button

A control that lets an operator stop an agent action before it executes.

Maker-Checker for AI

The two-person rule from banking, applied to agent actions: one entity acts, another approves.

Pre-Action Approval

A human gate that runs before a sensitive tool call, not after.

Compliance

How to express, version, and verify agent controls.

Policy as Code for AI

Declarative YAML rules in version control, reviewed in PRs, enforced at runtime.

Shadow Mode Validation

Recording policy decisions without enforcing them, so you can tune rules before enforcement.

Agent governance terms

New definitions for the vocabulary buyers use when agent actions need proof, policy, and approval.

Agent runtime governance

Agent runtime governance defined for AI agent teams building runtime authorization, approval, and audit controls.

Pre-action authorization

Pre-action authorization defined for AI agent teams building runtime authorization, approval, and audit controls.

Tool-call audit trail

Tool-call audit trail defined for AI agent teams building runtime authorization, approval, and audit controls.

Approval gates

Approval gates defined for AI agent teams building runtime authorization, approval, and audit controls.

Approval fatigue

Approval fatigue defined for AI agent teams building runtime authorization, approval, and audit controls.

Policy enforcement point

Policy enforcement point defined for AI agent teams building runtime authorization, approval, and audit controls.

Policy decision point

Policy decision point defined for AI agent teams building runtime authorization, approval, and audit controls.

Agent service account

Agent service account defined for AI agent teams building runtime authorization, approval, and audit controls.

Capability versus authority

Capability versus authority defined for AI agent teams building runtime authorization, approval, and audit controls.

Action-level evidence

Action-level evidence defined for AI agent teams building runtime authorization, approval, and audit controls.

Signed decision record

Signed decision record defined for AI agent teams building runtime authorization, approval, and audit controls.

Agent action ledger

Agent action ledger defined for AI agent teams building runtime authorization, approval, and audit controls.

Deterministic guardrails

Deterministic guardrails defined for AI agent teams building runtime authorization, approval, and audit controls.

Review-required action

Review-required action defined for AI agent teams building runtime authorization, approval, and audit controls.

Tool scope creep

Tool scope creep defined for AI agent teams building runtime authorization, approval, and audit controls.

Use these definitions at the tool boundary.

Sign upInstall the SDK