Human in the loop means you can see what a coding agent is doing, steer it, and stop it. For coding agents, that means visible steps, a way to redirect, limits on what the agent may do, and review before anything ships. Friday, the free open-source app, gives you each of these on your own computer.
What controls should you look for?
| Control | What it does |
|---|---|
| See every step | Shows what the agent is doing as it happens |
| Redirect | You send the agent a new instruction or a note on a line |
| Interrupt | You can stop a run without restarting it |
| Limits | Sets what the agent is allowed to do |
| Review | Lets you accept a change or send it back |
| Undo | Rolls a run back |
How do you set limits in Friday?
Each agent works in one of three modes: Plan (read and propose only), Manual (change files, and ask before running commands) or Auto (edit files and run commands without asking). Pick the mode per agent.
Every run ends in a change set you read before you keep it. Leave a note on a line or a range and send it back to the agent. Tasks on the shared task board are checked before they count as done: their tests run and another agent reviews them. Each agent can work in its own git worktree, so a run you reject leaves your own branch alone.
When should a person approve something?
- Anything that deletes or overwrites data.
- Anything that touches production.
- Anything that spends money or sends messages to other people.
What does human in the loop not do?
It does not make an agent correct. A person still has to look. See how to review AI-generated pull requests. Friday itself is free and MIT licensed; the download page has the installers.