Human-in-the-loop

What human-in-the-loop means for AI agents

A coding agent can take many steps of a task on its own: it reads files, changes them and runs commands. Human-in-the-loop does not mean that a person approves each of its steps. It means it is known in advance at which points a person makes the decision, for example:

• acceptance: a person judges the result and accepts it or sends it back for rework;
• scope: a person sets what the agent may do on its own and what it must ask about;
• questions: when the task leaves something open, the agent asks instead of guessing.

The point is that what matters for the result is not left to the model alone. Some of those decisions are made in the console itself: in the settings of Claude Code and Codex you set when the agent must ask for your permission before it changes a file or runs a command.

Human-in-the-loop in featureloom

In featureloom the agent takes a plan from the board, delivers it and reports it by itself. A person decides on acceptance, on rights and on the answers to the agent’s questions.

Acceptance belongs to a person. You review a delivered plan in the app, in the Acceptance section: you press Accept or Do not accept and can add your reasons. The server refuses a verdict sent from a console or an automation. The app records who decided and when, and your reasons if you give them. The verdict goes into the plan document too. After a rejection, the agent reads the verdict, takes the plan back and reports the rework.

Rights set what a console may do. A console, an agent connected to featureloom, acts with the rights of the person who connected it: it always reads, and it delivers, writes and deletes plans when that person has the right to do so, and not otherwise. An administrator grants rights on the People screen, and a change applies from the console’s next call. People set the column and the order of plans.

The agent asks when it is missing something. The Realizer prompt, which every new organization has from the start, tells the agent not to guess: when the plan leaves something open, the agent asks a question in the plan. The plan’s card then shows Question, and you answer on the plan’s screen with the Answer button. The question and the answer stay in the plan document.

Rights in featureloom cover what a console may do with plans and prompts. You decide what the agent may do on your computer in the console’s own settings, because featureloom does not connect to your code.