A coding agent spends most of its life waiting — on a build, on a test run, on a model. Watch one in a terminal and you spend that time waiting with it.
Run ten and the waiting overlaps.
What runs at the same time
Up to 10 Claude sessions at once, and up to 20 on Claude Max. Sessions are spread across your machines, so a single box is not the limit — a session is a conversation with an agent on a machine, and the console holds all of them in one tab.
Multiple Claude accounts are supported. A team is not one login, and the console does not make it behave like one.
It does not make one agent faster
Worth being precise about, because it is the thing people expect and it is not what happens. Nothing here speeds up a single agent: a build takes what a build takes.
What changes is how much of that time is yours. Ten agents working at once means ten builds running while you read the one diff that is ready. The gain is in the overlap, not in the model.
What makes it survivable
Running ten agents is easy. Running ten without losing track of them is the part that needs building:
- Nothing waits silently. A session that needs an answer surfaces in the rail rather than sitting behind a tab you have not clicked.
- Sessions on the same machine do not collide. Give each one its own git worktree and branch — see Worktree Isolation.
- Approvals stay readable. Ten agents produce ten times the approvals, which is exactly when a wall of raw tool calls stops working. See Approvals You Can Read.
- Everything is kept. Transcripts survive a restart, a reconnect, and a closed tab, so catching up on the session you left an hour ago is reading rather than archaeology.
The honest limit
The ceiling is not the console — it is you. Ten agents can generate more decisions than one person can make, and the answer to that is the approval tiering rather than more parallelism: read-only calls run without asking, writes ask once, and destructive calls are held. Turn that off and the number of sessions you can actually supervise drops sharply.