What is human in the loop?
Human in the loop means a person stays part of an automated process: they approve, correct or reject what the system proposes before it takes effect. For AI agents, it is the permission prompt before a command runs or the review before a publish. The point is to put a person where a mistake would be expensive.
Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.
The agent works on its own through the routine parts of a task: reading, searching, drafting a change.
Before an action that is hard to undo, it stops and shows a person what it is about to do: the command, the diff, the message.
The person approves, edits or rejects. Nothing reaches the real system until they do.
Approved patterns can be remembered, so the prompt returns only for new or riskier actions and the person's attention goes where it matters.
A record of who approved what keeps the decision reviewable later.
Human in the loop versus human on the loop
In the loop means the system waits. A person approves each consequential step before it happens, and if they walk away the work stops. On the loop means the system runs and a person watches, able to intervene but not asked to. Out of the loop is full autonomy. The terms predate AI, and the useful part is that they describe where the checkpoint sits, not whether a person is involved at all. Code review is in the loop. A dashboard with an alert is on the loop.
Example: a permission prompt in a coding agent
Claude Code is a clear case. In its default mode it asks before running a shell command or editing a file, shows exactly what it is about to do, and offers yes, no, or yes and stop asking for this pattern. Plan mode goes further: Claude reads and explores but does not edit your source files until you switch modes. At the other end, a bypass mode skips the prompts, and the documentation says to use it only in an isolated environment such as a container where the agent cannot cause damage. Each mode is a decision about where the human sits.
The MCP specification takes the same position for tools in general. It says there should always be a human in the loop with the ability to deny a tool invocation, and that clients should show a tool's inputs to the user before calling the server.
Why it matters
Agents fail in two ways: they misunderstand you, and they get misled by what they read. A person at the right checkpoint catches both. OWASP's guidance on prompt injection lists human approval for high-risk actions as a mitigation precisely because the model-side defenses are probabilistic and the approval step is not. The agent can be fooled into wanting to delete a directory; it cannot be fooled into having deleted it if a person said no.
The loop does not have to be a modal prompt. On an Intraface board, an agent publishes a new version, every version is kept with who made it, and anyone can pin a comment to the page addressed to the agent. The person is in the loop after the fact rather than before, which is the right trade when a publish is cheap to undo and a prompt would interrupt five teammates' agents.
Common mistakes
- Approval fatigue. A prompt for every file read trains people to press yes without looking. Auto-approve the reversible and cheap; reserve the prompt for what is not.
- The checkpoint in the wrong place. Approving the plan and then letting execution run unattended protects against misunderstanding, not against an injection that arrives mid-task.
- Mistaking instructions for enforcement. A line in a rules file asking the agent to check with you is a request. A permission rule or a hook that blocks the call is enforcement. Claude Code's docs draw exactly this distinction.
- No record. If nobody can see later who approved what, the loop existed but taught you nothing.
FAQ
- What does human in the loop mean for AI agents?
- A person approves, edits or rejects the agent's consequential actions before they happen: running a command, editing a file, sending a message, publishing a page. The agent does the routine work and stops at the checkpoints you set.
- Does keeping a human in the loop slow an agent down?
- At the checkpoints, yes, and that is the point. The cost stays small if prompts are reserved for actions that are hard to undo and approved patterns are remembered. Prompting for everything is what makes people stop reading and press yes.
- Where does the MCP spec say there should be a human in the loop?
- In the tools specification. It says there should always be a human in the loop able to deny tool invocations, and that applications should show which tools are exposed, signal when one is called, and present confirmation prompts.
Stop sending HTMLover Slack.
Create an account in a minute, connect your coding agent, and make your first board.