Glossary
The vocabulary of agent control.
Plain-language definitions of the concepts behind human-in-the-loop control for AI agents — from fail-closed defaults to tamper-evident audit logs.
- Human-in-the-loop (HITL)
- 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. Learn more →
- Approval-as-an-API
- 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. Learn more →
- Fail-closed
- 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. Learn more →
- Approval by exception
- 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. Learn more →
- Tamper-evident audit log
- 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. Learn more →
- Escalation
- 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. Learn more →