Learn · AI agent with approvals
What are AI agent approvals?
By Lalla Asmaa Alaoui, founder of Mel · Updated
AI agent approvals decide which of an agent's commands and file edits need your yes. They are the dial between an agent that can do nothing without you and one that can do anything.
Approvals are what make agents usable on real work. An agent that can do nothing without you is a chat. An agent that can do anything without you is a risk. The best setting depends on the project, the task and how much you trust the agent yet.
What should need an approval?
A practical rule: approve what is hard to undo, or reaches outside the project.
- Usually approve: deleting or moving files,
git push, deploys, database migrations, anything withsudo, installing packages, edits to config, CI or secrets files. - Often safe to allow: reading files, listing folders, viewing
git statusandgit diff. Even these can show secrets, so the agent should not print key files. - Depends on you: editing source files inside the project. Many people approve edits at first, then relax this once they review diffs as a habit.
Good approvals are configurable
One fixed policy never fits. A scratch prototype and a payment service need different rules. Look for:
- Separate controls for commands and edits. They carry different risks.
- Per-project settings. Strict where it matters, relaxed where it does not.
- Clear cards. An approval should show exactly what will run or change.
- Easy changes. If changing a rule takes a config file and a restart, people stop adjusting it.
How Mel does it
- Commands and edits are both covered. You configure which of them need your yes.
- One set of rules for the whole team of agents. Every agent on the task works under it.
- Loosen as trust grows. Start strict, then relax the routine actions you keep approving.
- File work counts too. Moving or renaming files runs as commands, under the same approvals.
How the approaches compare
| No approvals | Approve everything, always | Configurable approvals (Mel) | |
|---|---|---|---|
| Speed on routine steps | Fast | Slow, many interruptions | Fast where you allow it |
| Risk on destructive steps | High | Low | Low where you require approval |
| Fits different projects | No | No | Yes |
| Review fatigue | None, but no control | High, people click yes without reading | Lower, approvals are where they matter |
| Good for | Throwaway sandboxes | Your first hour with a new tool | Daily real work |
Habits that make approvals work
- Read the command. The approval card is only as good as your attention.
- Start strict, then loosen. Relax rules for actions you have approved many times without surprise.
- Keep git clean before a task. Then the diff shows exactly what the agent did.
- Review the diff at the end, even for edits you allowed.
FAQ
What is an AI agent with approvals? An agent that can act on your computer but needs your yes for the commands and edits you choose to control.
Do approvals slow agents down? Only on the actions you choose to control. Allow the safe, routine ones and the agent keeps moving.
Can I approve edits but control commands differently? Mel's approvals cover both commands and edits, and you configure which of them need your yes.
Is an agent with approvals safe for production repos? It is safer than one without. Combine strict approvals with code review and your normal branch protection.
What if I approve something by mistake? Work in version control. A clean starting tree means any change can be reviewed and reverted.
Mel – the agentic terminal · openmel.dev. A multi-agent, multi-model AI terminal for Mac, Windows, Linux and VS Code. Free with your own API key.
Lalla is a hands-on tech founder, agentic infrastructure architect and the vision behind Mel.
Follow on LinkedIn