AI Concept Atlas
FIELD NOTES / 01
Guided learning · 6 lessons

Making an agent dependable

Separate testing, execution evidence, access boundaries, and human decisions.

Lesson 6 of 6

Human-in-the-loop

Some decisions also need a person’s review before execution continues.

Human-in-the-loop introduces a person’s decision into an automated process. In an agent workflow, execution can pause at a review point so a person can approve, change, or reject a proposed action.

[LangChain]

In practice

An assistant drafts a customer email and waits for an operator to approve or edit it before sending.

[LangChain]

A little deeper

The system needs a defined review point and a way to continue from the decision. The presence of a person somewhere in the process does not imply that every action receives review. [LangChain]

A common mix-up

Human oversight means reviewing every model token.

Review can target selected actions or decisions. [LangChain]

Helpful to know: Agent

Explore Human-in-the-loop and its connections ↗

← Previous lesson