← Back to Concept Contrast

Concept Contrast Fable

Human-in-the-loop and approval flows: one is real-time correction, the other is sign-off in advance

Conclusion first:Every Agent system needs a "human" added somewhere, but where and what they do is often glossed over in the plan. Human-in-the-loop means the model continuously hands key decision points to a person during operation, and the person can interrupt or correct at any time; an approval flow adds a sign-off gate after the model produces results and before release. Different positions mean completely different costs, latency and risk windows.

Every Agent system needs a "human" added somewhere, but where and what they do is often glossed over in the plan. Human-in-the-loop means the model continuously hands key decision points to a person during operation, and the person can interrupt or correct at any time; an approval flow adds a sign-off gate after the model produces results and before release. Different positions mean completely different costs, latency and risk windows.

Fable

Two different proofreading methods in a printing house

The old print shop printed large numbers of advertising flyers every day. In earlier years, after the typesetter finished a full page, it was handed to a proofreader to read from beginning to end, sign after finishing, and then go to press. This process had been used for more than ten years, and everyone felt it was naturally how things should be; slow was just slow, and at worst an error meant reprinting.

Later they took on a batch of pharmaceutical instructions, where mistakes could cause accidents. The typesetter proposed trying a new method: after setting each key passage—dosage, contraindications, usage—he pushed that small passage to the proofreader's screen, and the proofreader confirmed or sent it back for revision on the spot. Typesetting and proofreading proceeded in alternation, and by the time the full page was set, the key passages had already been confirmed over multiple rounds.

Printing efficiency actually went up. Previously, proofreading happened after the whole page was set, and finding a typo in the headline meant taking apart and resetting most of the page. Now errors are caught at the paragraph level, and the typesetter does not have to go back and dismantle a page that is already locked in.

After the new supervisor took over, he felt the two methods were too complicated and unified them into "all documents must be signed off by a proofreader before printing." He removed paragraph-by-paragraph proofreading for pharmaceutical instructions and replaced it with a one-time review after layout was complete. An accident occurred in the third week: a disclaimer was set two point sizes too small, no one checked it during layout, and by the time the proofreader reached that line, the plates had already gone to press.

Only in the retrospective did everyone clearly see the fundamental difference between the two kinds of sign-off. Section-by-section proofing embedded human judgment into the typesetting process—at every step the typesetter could be stopped. Human in the loop. One-time review confirmed the final draft only after the entire typesetting process was finished—an approval flow. The latter can only block the final result; it cannot block hidden risks in the intermediate steps.

The factory later clearly separated the two types of processes: high-risk materials used section-by-section review, where a person could halt at every step; low-risk printed materials used full-plate final review, with only one signature gate. Process cards and shift schedules clearly marked which category each check belonged to.

The most unexpected gain was the scheduling data. Although section-by-section proofreading involved many steps, the overall completion time was actually shorter—because errors were eliminated during production and did not require full-page resetting. Final approval signing looked simple, but a single error hidden at the bottom layer could disrupt the delivery date for several days.

Human in the loop

During model operation, key decision nodes are handed to humans for real-time judgment, correction or interruption, and human feedback continuously influences subsequent output.

Approval flow

After the model completes all output, a human reviews and signs off before it can be published or executed; this is an ex-post control gate.

Story Mapping

  • Section-by-section typesetting pushed instantly to the proofreader: human in the loop, continuous review during operation
  • Sign off once after the full version is laid out: approval workflow, post-hoc gate-style review
  • New supervisor unified everything into final review, causing an incident: mixing two modes, high-risk scenarios lost intermediate interception
  • Two types of materials use two types of processes: choose HITL or an approval flow based on risk level

Applying it to an Agent project

When designing human-agent collaboration, classify first: which steps require a human to stop things midway, and which only require sign-off afterward. Those requiring midway stops need interaction interfaces and interruption protocols designed, which is costly but narrows the risk window; those requiring only sign-off afterward must ensure outputs are auditable and traceable, which is low-cost but means bearing all the risk of the intermediate steps.

京ICP备2026014272号