How to run multiple AI agents at once without losing track
Run several Claude Code agents at once and still know which is blocked, which finished, and which is burning budget.
To run multiple AI agents at once, give each one its own terminal or worktree so they do not fight over the same files, then connect them to one control panel so every session reports to a single screen. The hard part is not starting three agents. It is knowing, at a glance, which one is blocked on your answer, which one finished, and which one is spending money while you look the other way.
Key takeaways
- Isolate each agent (separate terminal, separate git worktree) so concurrent edits do not collide.
- Route every session to one board so you can tell "blocked" from "running" from "done" without reading three terminals.
- Let permission policies and a kill switch cover the agents you are not watching, instead of babysitting each prompt.
Start with isolation, not just more terminals
Two agents editing the same files at the same time will overwrite each other and leave you with a merge you did not ask for. Before you scale up, give each agent a clean lane.
The simplest version is one terminal per agent. The cleaner version is one git worktree per agent, so each session has its own checkout of the branch it is working on and they never touch the same working tree. Name the branches after the task. You can mix agents too: a Claude Code session refactoring one module, a Codex job writing tests, a Gemini CLI session on docs. They run in parallel because they are not in each other's way.
That solves the file problem. It does nothing for the attention problem.
The real pain: which one needs you right now
Run three or four agents and the failure mode shows up fast. Your terminals scroll past faster than you can read. One agent is sitting on a permission prompt waiting for a yes. Another finished ten minutes ago and is idle. A third is still churning, and you have no idea how much it has spent. A plain notification stream makes this worse. Every event reads as "agent needs input" with no way to tell a git push approval from a finished build.
You end up with two bad habits. Mute everything and miss the approval that mattered, or leave it loud and start tapping through prompts without reading them. Tapping through is how an agent runs a command you would never have allowed.
Put every session on one board
The fix is to stop watching terminals and watch one screen instead. The Fleet board puts every connected agent on a single view, in lifecycle columns, so the three questions you actually care about have answers:
- Which agent and what task. Each card carries the agent name and a real task title captured from what you asked it to do, so you see "refactor the billing webhook" instead of
session 4f2a. - What stage. Cards sit in columns by lifecycle: waiting on you, still running, finished. Blocked sessions stop hiding inside terminal scrollback.
- Is this one session or five. If an agent re-prompts or fires the same question twice on a retry, the board folds those into one entry instead of making it look like your whole fleet caught fire.
From the board you approve, deny, or send a correction back to the running agent. You answer from your phone too, including from the lock screen with the app closed, so you do not have to be at the desk to unblock a session.
Notifications are routed by presence. In Smart mode, if you are at the terminal the question stays local and your phone stays quiet. Step away and the next approval lands on your phone. Routing only changes where a notice goes, never whether an action is allowed.
Let policy and a kill switch cover the agents you are not watching
Watching a board is still watching. The point of running many agents is to not babysit each one, so push the routine decisions into rules.
Permission policies match on the actual command, not just the tool name, so you can allow git status and still gate git push. A read-only safe floor auto-approves proven read-only shell commands like cd, ls, cat, and git log, a default that came out of looking at 1,721 real production questions. Permission Autopilot mines your own approve and deny history into one-tap rule suggestions, so the more you run, the fewer prompts you see. Every decision lands in the audit trail with what changed and where you answered from, which is the receipt you want when four agents have been working unsupervised.
When an agent does go off the rails, you do not want to dig through three terminals to stop it. The kill switch ends a running session from the board or your phone, so a job that is looping or heading somewhere you did not intend gets cut off fast instead of churning while you are looking at the other three.
Honest caveats
This works for any agent connected through the CLI hook: Claude Code, Codex, Gemini CLI, Cursor, and Hermes. Claude Desktop connects by pasting one URL, but it has no hooks, so that connector can only notify and ask, not enforce a gate. On an iOS home-screen install the cross-origin deep link is broken, so iOS surfaces pending questions through an inbox rather than a tap-through link. Android and the native app handle the link directly.
The workflow is the same whether you run two agents or ten: isolate them, route them to one board, and let policy plus a kill switch cover the ones you are not actively watching. Connect your first agent with the quickstart, wire up Claude Code notifications, and see the full setup on the agents overview. It comes with every paid plan, $9.99 a month for Agent and $19.99 for Agent Pro, on pricing.