Human-in-the-Loop

A workflow where an AI agent does the work but a human approves it before it counts: the agent proposes, you decide. The standard pattern for mid-trust agents.

All Dictionary

What it is

Human-in-the-loop (HITL) means an agent’s work does not take effect until a person approves it. The agent drafts the token fix, opens the pull request, or proposes the component change, and a human reviews and merges. The agent does the labor; the human keeps the judgment.

Why this matters for designers

HITL is where most design system agents should live for a long time. It captures most of the speed (the agent does the tedious 90%) while keeping the failure mode survivable (a bad proposal gets rejected, not shipped). It also generates the evidence you need for trust-level promotions: every approved or rejected proposal is a data point about whether the agent deserves more autonomy.

How it works in practice

  1. The agent completes a task end-to-end but cannot make it final: its output lands as a proposal (a PR, a draft, a report).
  2. A human reviews with a clear rubric: would I have shipped this?
  3. Approvals and rejections are tracked; they become the agent’s track record.
  4. As the record improves, review can loosen, from every change to spot checks, which is exactly a trust-level promotion.