Human-in-the-Loop (HITL)
Human-in-the-loop (HITL) is a control pattern where an autonomous AI agent pauses for explicit human approval before executing a high-risk action, so a person reviews and authorizes the operation before any side effect occurs.
Definition
Human-in-the-loop (HITL) is a control pattern where an autonomous AI agent pauses for explicit human approval before executing a high-risk action, so a person reviews and authorizes the operation before any side effect occurs.
Approval as a governance decision
Not every action should be fully autonomous. Sending money, deleting data, deploying to production, or messaging customers are operations where a human checkpoint is appropriate. In an agent control plane, human-in-the-loop is expressed as a policy decision: when a rule matches a high-risk action, the Safety Kernel returns REQUIRE_APPROVAL and the job is held in an approval queue instead of executing. A reviewer sees the agent's intended action and context, then approves or rejects it.
Making approvals reliable
Approval flows must be robust under real-world conditions. Cordum binds an approval to the specific job and evaluates it against a policy snapshot, so a decision cannot drift if rules change mid-review, and it guards against double-approval and lock contention so the same action is not executed twice. Done well, HITL gives teams a graduated autonomy model: low-risk actions run freely, high-risk ones wait for a human, and every approval is recorded in the audit trail.
Frequently asked questions
When should an agent action require human approval?
Reserve approvals for actions whose blast radius justifies the latency — irreversible or high-impact operations like payments, deletions, production changes, or external communications. Routine, low-risk actions should run autonomously so the approval queue stays meaningful.
How is human-in-the-loop enforced in a control plane?
It is a policy outcome: a matching rule returns a REQUIRE_APPROVAL decision, and the control plane holds the job in an approval queue until a reviewer acts. The decision and its outcome are written to the audit trail.
Related reading
Govern your AI agents with Cordum
Cordum is the agent control plane: policy-before-dispatch enforcement, human approvals, and a tamper-evident audit trail for autonomous AI agents.