Question 1

Human-in-the-loop (HITL)

Accepted Answer

A control pattern where a person must approve or reject a system's action before it takes effect. For AI agents, it means pausing an irreversible action until a human decides, rather than letting the agent act unchecked.

Question 2

Approval-as-an-API

Accepted Answer

A hosted service that adds a human approval step to any action through a single API or SDK call — handling routing to a reviewer, verdict enforcement, and audit logging so you don't build them yourself.

Question 3

Fail-closed

Accepted Answer

A safety default where any unknown, missing, or malformed input resolves to the restrictive outcome. In an approval gate, uncertainty means the action does not run — the permissive path is never the default.

Question 4

Approval by exception

Accepted Answer

A policy model where most actions proceed automatically and only exceptions — high amounts, sensitive scopes, risky agents — are routed to a human. It keeps agents fast while still guarding what matters.

Question 5

Tamper-evident audit log

Accepted Answer

An append-only record where each entry is cryptographically hash-chained to the previous one. Any modification to a past entry breaks the chain and is detectable, so the history of decisions is verifiable.

Question 6

Escalation

Accepted Answer

Routing an approval request to a higher authority or a wider reviewer pool when it isn't handled in time or exceeds a risk threshold — ensuring high-stakes actions don't stall or slip through.

The vocabulary of agent control.