What is human-in-the-loop? Integrating human review into AI execution
Understand human involvement through an outgoing quotation draft, including observation, edits, approval, pause-resume behavior, and preventing repeated execution.
Understand the concept
Human-in-the-loop adds human judgment, edits, or approval at defined points, such as reviewing an external action. It can refer to training annotation or runtime oversight. A confirmation button alone is insufficient: execution must pause and approval must bind to the specific action that follows.
Show reviewers a concrete action they can assess
For an outgoing quotation, show recipients, full content, and attachments. A generic continue prompt gives insufficient context. If the model changes the content after approval, the original approval may not cover it.
This is a runtime review design without sending a quotation. People can edit or reject, and the program must record that decision instead of treating any reply as assent. Required review depends on the operation, not only the model’s risk rating.
Bind pauses and resumption to the original task state
LangGraph provides interrupt-based pauses and resumption from saved state. Developers must understand restart points, especially whether side effects before a pause can run again. The interface is only part of the mechanism; the runtime enforces blocking and resumption.
If input changes during review, show the difference and verify what is being approved. Bind identity and session so one person’s confirmation cannot resume another action. Google ADK discusses oversight and action controls, while applications still integrate business authorization.
Validate oversight through execution evidence
In a test environment, verify that rejection causes no external action, approval executes once, and edits determine the executed version. These are proposed checks without business-system actions. Click counts alone do not establish effective blocking; inspect state and execution records.
Reviewers can tire or err, so provide evidence and avoid attention-draining prompts for every tiny step. Different-impact actions can use different policies, with controlled bypass paths. Oversight is a process design, not a disclaimer for AI errors.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.