How to manage AI coding agents

To manage AI coding agents well, you need four things: a view of what every agent is doing, a signal when one needs you, rules for what each may run, and a clear record of what each changed. Without them, more agents just means more tabs to check.

Mel – the agentic terminal gives you all four in the same window where the agents work. One agent is easy to watch. Three agents in three tabs, one of them a Claude Code session and one a Mel team, is a management job.

The four parts of managing agents

1. See what each agent is doing. Which agents are running, which are idle, which finished. A list you can read at a glance, not a row of terminal windows.

2. Know when an agent needs you. An agent waiting for an approval or an answer does nothing until you come back. The signal has to reach you wherever you are looking.

3. Set what each agent may do. Which commands and edits run on their own and which need your yes. A step that cannot be undone should always ask.

4. Review what changed. The diff, not the summary. And who changed it, when several agents worked on one task.

How Mel does it

• The Runs screen. One screen lists every agent run on your computer in three groups: Needs you, Running and Recent. Open it from the titlebar or with Ctrl+Shift+M. A click takes you to that agent's pane.
• A status in the sidebar. Vertical tabs show every tab and pane with a live status: running or blocked.
• Notifications. When an agent in another tab is blocked on an approval or finishes its turn, Mel raises a card and adds it to the inbox, with an unread count.
• CLI agents that went quiet. When Claude Code or another CLI agent stops printing for a few seconds and you have not looked at its pane, Mel treats it as waiting for you, once.
• Worker pills for a team. While a team of agents runs, a strip shows the lead and each worker with its status, model and tokens.
• Configurable approvals. Decide which commands and edits need your yes. A command that cannot be undone always asks a person.
• The review panel shows every change, with line comments and an editor.

A simple routine

1. Start each agent with a finish line, such as "npm test passes".
2. Keep approvals on for anything that deletes, deploys or touches production.
3. Work from the Runs screen: clear Needs you first, then check Running.
4. Review each diff before you commit, one agent at a time.
5. Check usage after the first few runs, so you know what a typical task costs.

FAQ

How do I monitor several AI coding agents? Use one place that lists every run and its state. In Mel that is the Runs screen, plus the sidebar status and the notification inbox.

How do I know when an agent is waiting for me? Mel marks the session blocked, raises a notification and lists the run under Needs you.

Can I control what an agent is allowed to run? Yes. Mel's configurable approvals decide which commands and edits need your yes.

Does this work with Claude Code and Codex? Yes. They run in Mel's panes and show in the sidebar, and when one goes quiet waiting for you Mel lists it under Needs you.

• How to run multiple AI coding agents in parallel
• What are AI agent approvals?
• How to review an AI agent's code changes
• What is a terminal for AI coding agents?