How to run multiple Claude Code agents at once

One agent is useful. A fleet of them is a different job entirely. The moment you go from a single Claude Code session to five running in parallel, the bottleneck stops being the model and starts being you: your attention, your terminal tabs, and your ability to remember which agent was doing what.

This is the problem SanuDesk was built to solve. Here's how to run several coding agents at once without drowning.

Give every session a home

The naïve approach is a pile of terminal tabs. It falls apart fast: tabs look identical, you lose scrollback when something crashes, and there's no sense of "which project is this?"

Instead, tile your sessions into a grid. Spawn two, four, or six agents and let them auto-arrange so you can watch all of them at a glance. Maximize one to focus, then drop back to the grid. Each session belongs to a workspace (one folder, one project), so you always know where an agent is working.

Let the board hand out the work

Watching agents is only half of it. The harder part is deciding what each one should do next. A Kanban board turns that into a queue: write a card with a title, some Markdown notes, and acceptance criteria, then deploy it. The card becomes the prompt for a fresh agent session, moves to In Progress, and lands in In Review, with a "what changed / how to test" note, the moment the agent stops.

Now you're not typing prompts into ten terminals. You're grooming a backlog and approving results. Scope each card by the files it touches, not just the outcome it delivers: it's the cheapest way to stop two agents colliding in the merge an hour later.

Mix engines on purpose

Different work suits different engines. You might send a gnarly refactor to a Claude agent, a quick script to Codex, and a research spike to Gemini, each on its own card, each remembering your effort level and permission flags for next time. Running multiple engines side by side is less about loyalty to one model and more about routing each task to the tool that does it best. The same routing instinct applies to where an agent runs: a chore that needs no steering can head to a background agent in the cloud, while anything you want to watch stays in a session you can see.

Keep the fleet from going stale

The failure mode of a big fleet is silent stalls: an agent finishes, or gets stuck waiting for input, and you don't notice for twenty minutes. Two habits fix this:

• Supervise over MCP. Have one agent watch and steer another's terminal: read any pane's output and type into a live session, so the fleet can partly manage itself.
• Automate the recurring stuff. Anything you run on a schedule ("triage new issues", "update the changelog") becomes a Loop that fires on its own and journals every run.

The short version

Running many agents well comes down to three things: see them all (a tiled grid), feed them from one place (a board), and don't babysit (supervision and automation). Get those right and a fleet of agents feels less like ten things to manage and more like one system that ships.

Want the desktop app that does all of this? Download SanuDesk free and open your first project.