Glossary
Human in the loop
A design pattern in which an AI agent pauses for explicit human approval on consequential actions — sends, calendar writes, destructive operations — instead of acting fully autonomously.
Human in the loop (HITL) is the deliberate insertion of a person into an agent's action pipeline before externally-visible side effects fire. The framing matters because the alternative is a binary: either the agent is fully autonomous (scary, hard to undo) or the user reviews every output (tedious, leads to rubber-stamping). HITL is the middle band — autonomy on low-risk read work, approval gates on high-risk writes.
Helix's approval policy is the HITL primitive. Three trust levels — "always", "sensitive", "never" — express the global posture, and per-identity rules sharpen them by recipient, domain, subject, time window, attachment, and first-time conversations. A common shape is "always for board members, sensitive for clients, never for internal teammates," which is closer to how humans actually delegate than a single global slider.
The well-known HITL failure mode is habituation: when a queue is too noisy, approvers click "approve" without reading. Helix mitigates this in two ways. First, sensible defaults — read actions never queue, write actions always do unless explicitly downgraded. Second, the approval queue surfaces the agent's reasoning and the matched rule, so the approver sees why the agent thinks the action is correct.
Related terms
In product