What it is
Human-in-the-loop (HITL) is a control mechanism where ambiguous or high-severity agent actions are paused and escalated to human review before execution. The agent cannot continue until a human approves, denies, or modifies the action.
HITL sits between pure automation (agent makes all decisions) and pure manual (human makes all decisions). It preserves agent autonomy for routine tasks while enforcing human oversight for consequential decisions.
Why it matters
Agents are non-deterministic. An agent might attempt an action that isn't clearly allowed or clearly denied by policy rules. It might be attempting something legitimate in an unusual way, or it might be attempting something harmful. A human reviewer can make that judgment faster than encoding it into increasingly complex policy rules.
HITL also provides accountability. If an agent deletes important data, the audit trail shows whether the action was human-approved or auto-allowed. Regulators and auditors care about that distinction.
How it works
When the proxy evaluates a tool call against policy and determines that the action requires escalation (either because it matches a medium-severity pattern or because the policy is configured to escalate that type of action), the proxy pauses the tool call and notifies the human-in-the-loop system.
A notification is sent to the SOC team with context: agent ID, action attempted, reason for escalation, timestamp, severity. The human reviewer examines the context and makes a decision: approve, deny, or modify the action. The decision is logged in the audit trail and returned to the agent as a structured response.
How Intercis implements it
Today, Intercis's policy engine returns a verdict on every tool call on the LLM API wire — allow, deny, or observe — from a deterministic deny-list of 110 regex patterns plus an LLM classifier. It does not run a live human-in-the-loop approval loop: pending exists as a verdict state for actions that would require human review, but no reviewer approve/deny action is wired up yet, and nothing auto-terminates a session on a policy match.
Because enforcement is deterministic rather than routed to human reviewers, OWASP T10 (Overwhelming Human-in-the-Loop) does not arise here: a matched deny policy blocks the call outright rather than queuing it for a person to approve.
Related terms
- AI Agent Kill Switch — HITL can trigger termination.
- AI Agent Policy Enforcement — HITL is the human-review (escalate) decision; Intercis's live verdicts are allow, deny, and observe.