The gate spectrum
Lesson 1 of 5 in Human-in-the-Loop: Designing the Human Back In.
In foundations you learned the autonomy pyramid: fully manual, suggest, act-with-approval, act-and-report, fully autonomous — a dial set per action, not per agent. This module is about engineering the human side of that dial. Because “put a human in the loop” is not one design. It is at least three, and picking the wrong one is how teams end up with oversight that exists on the org chart but not in reality.
The three canonical gate shapes differ on exactly one question: when does human judgment enter — before the action, after it, or inside it?
- Notify-after. The agent acts; a human finds out. The gate is really an audit: digests, alerts, traces someone actually reads. This is the act-and-report tier — often called human-on-the-loop rather than in it.
- Approve-before. The agent prepares the exact action and stops. A human sees it and clicks approve or deny before anything fires. This is the act-with-approval tier — the approval gate proper, and the shape most people mean by human-in-the-loop.
- Co-draft. The agent produces a draft; a human edits it and performs the action themselves. This is the suggest tier. The human is not reviewing the agent’s action — the action is the human’s, with agent-supplied raw material.
| Gate shape | Pyramid tier | Who performs the action | What the human sees | What a wrong action costs |
|---|---|---|---|---|
Co-draft | Suggest | The human, using the agent’s draft | The full draft, editable, before anything happens | Minutes of wasted drafting — the human still owns the send |
Approve-before | Act with approval | The agent, after a human approves | The exact pending action, before it fires | One bad approval — the gate is the last line of defense |
Notify-after | Act and report | The agent, immediately | A record of what already happened | Whatever happened between the action and someone reading the digest |
Notice what the mapping buys you: the gate shape is not a philosophy debate, it is a position on the dial you already know how to set. An action’s reversibility, blast radius, and observability told you which pyramid tier it belongs on; the tier tells you which gate shape to build. A coding agent reads files ungated, edits its own branch notify-after, and runs shell commands approve-before — three gate shapes inside one agent, and all three are correct.
One more distinction matters before you sort anything: a gate is not a permission. Tool scoping decides what the agent can do at all; gates decide which of the permitted actions need a human decision in the moment. If an action should never happen, revoke the capability — don’t gate it and hope the human always says no.
Key terms: human-in-the-loop, approval gate, human-on-the-loop, autonomy, blast radius
Interactive sorting exercise: One agent, six actions. Pick the right gate shape for each — reversibility, blast radius, and who should own the final act are your criteria.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.