Glossary

The words used across the product and these docs, from agent key to two-person approval, each in a sentence or two.

Action
The name an agent gives a kind of thing it wants to do, such as email.send or payment.refund. Rules match on it, either exactly or by a start such as email.*. More on Action
Agent
A program that uses an AI model to act on someone's behalf: sending email, spending money, changing files or settings.
Agent key
The secret an agent sends with every call, starting with gl_. Each agent gets its own. It is shown once, stored only as a hash and can be revoked at any time. It can ask, and read, withdraw or report on its own requests; it can never list or decide requests. More on Agent key
allowed
The field in every answer that says whether the agent may act. It is true only after an explicit approval, and never while the workspace or the agent is paused. Pending, denied, expired and withdrawn all leave it false. More on allowed
Amount limit
A limit on a rule. An always-allow rule applies only up to it; an always-ask or never-allow rule applies above it, and also when a request states no amount or another currency. More on Amount limit
Approval fatigue
What happens when people are asked to approve too much: they stop reading and click yes. Rules keep the questions few enough to be read. More on Approval fatigue
Approved with changes
An approval of a narrower version than the agent asked: new wording, a lower amount, or both. The agent may then do only the approved version, given in approved_summary and approved_amount. More on Approved with changes
Audit trail
A permanent list of who asked for what, who decided, when and why, and how it went. In this product it is called the record. More on Audit trail
Budget
The most an always-allow rule may approve in total per day, week or month. Past it, requests go to a person until the period starts again. More on Budget
Decision
A person's approval or denial of a request, or a rule's. It is final: nobody can change it afterwards.
Emergency stop
One switch that pauses every agent in a workspace. While it is on, nothing is allowed, whatever was approved. Any member can pause; only an owner can resume. One agent key can be paused the same way. More on Emergency stop
Escalation
When a request that a rule sent to named people waits longer than the rule allows, anyone in the workspace may decide it, and a notice goes to the workspace's chat channel. More on Escalation
Excessive agency
OWASP's name (LLM06:2025) for an LLM system that can take damaging actions on bad or manipulated model output, through too many tools, too much permission or too much autonomy. More on Excessive agency
Expired
A request whose time limit ran out before anyone decided. It can no longer be approved, and the agent must not act.
Human in the loop
Designing an agent so that a person makes certain decisions before it acts, instead of reviewing what it did afterwards. More on Human in the loop
MCP
The Model Context Protocol, an open standard for giving AI agents tools. The approvals service is an MCP server with tools such as request_approval and check_approval. More on MCP
Outcome
What an agent reports after an approved action: succeeded or failed, with a short note. It completes the record. More on Outcome
Owner and member
A person who can manage a workspace: keys, rules, people and notifications. A workspace always keeps at least one. A member can decide requests and pause, but not manage. More on Owner and member
Pending
A request still waiting for a decision. The agent must wait or give up; it must not act.
PreToolUse hook
A command Claude Code or Codex runs before a tool call. The approvals hook asks a person and lets the call through only on a yes. More on PreToolUse hook
Remembered answer
A rule an owner writes from the inbox by approving or denying for the next hour, the next day or always. It covers exactly that action from exactly that agent key. More on Remembered answer
Request
One question from an agent: may I do this? It names the agent, the action and a summary for the person deciding, and has a time limit. More on Request
Risk tier
Which of let it run, ask one person, ask two people or never an action belongs in, judged by how hard it is to undo, whom it reaches, how large it is and who can steer it. More on Risk tier
Routing
An always-ask rule that names the people who decide what it catches. Until the request escalates, only they and the workspace owners can decide it. More on Routing
Rule
An answer given in advance by a workspace owner: always allow, never allow or always ask, for one action or a family of them. When rules disagree, the most cautious wins. More on Rule
Summary
The sentence the person deciding reads first. It should say exactly what will happen, with the amount, the recipient or the number of things affected. More on Summary
Time limit
How long a request waits for a decision: 10 seconds to 7 days, 15 minutes unless the agent says otherwise.
Two-person approval
An always-ask rule that needs two or three different people to approve. One denial stops the request at once. More on Two-person approval
Withdrawn
A request the agent took back before anyone decided, because its plan changed.
Workspace
The space that holds a team's agents, keys, rules, people and record. Nothing in one workspace can be seen from another. More on Workspace