Human-in-the-Loop Gate
Also: HITL · approval gate · human checkpoint · confirmation step
A required human approval checkpoint inserted before an agent takes a consequential or irreversible action.
Why CCAR-P tests this
CCAR-P tests placing approval gates on high-impact actions and enforcing them in code rather than relying on the model to ask.
A human-in-the-loop gate pauses the agent and requires an explicit human decision before a sensitive action proceeds — sending funds, deleting data, emailing a customer, deploying code. The agent proposes; a person approves, edits, or rejects.
The gate must be a deterministic control in code, not an instruction in the prompt. A prompt saying 'ask before deleting' can be overridden by a confident model or an adversarial input; a code path that blocks the tool call until an approval token is present cannot. This is the difference between a suggestion and a guarantee.
Gate placement is a design decision: gate the irreversible and the high-blast-radius, and let low-risk, easily-reversible actions run autonomously so the human isn't drowned in trivial approvals. Escalation should trigger on policy (this action class requires review) rather than on the model's self-reported confidence.
Exam trap
Believing a prompt instruction to 'ask the user first' is a real safeguard — a must-hold gate belongs in deterministic code, and escalation should key off policy, not the model's own confidence.
Commonly confused with
Related concepts
Independent, unofficial study material from Claude Cert Prep. Not affiliated with Anthropic.