How many AI agents can you actually run in parallel

Short answer: technically, as many as the machine holds, and on a modern Mac that is past ten. In practice the limit is human and arrives much sooner: with no interface help at all, two or three; with visible state per panel and alerts when one gets stuck, six to eight. Beyond that the bottleneck stops being following them and becomes reviewing what they delivered.

The technical limit, and why it is rarely the problem

Each AI agent is a terminal process plus a copy of the project on disk, if you isolate with worktrees. It uses little CPU while waiting on the model, which is most of the time. Disk is the real cost, and even that is modest: ten copies of a project without installed dependencies rarely pass a few gigabytes.

So the machine is not the brake. Whoever asks "how many can it handle?" is almost always asking the wrong question.

The real limit: how many you can follow

Following means knowing, at any moment, what each is doing and which one needs you. Reading each terminal, that saturates at two or three: past that you spend more time scanning text to rebuild in your head a state the machine already knows than actually reviewing work.

The maths that matters: the gain from running six agents is capped by how many you can follow. If you follow two, opening six produces the output of two plus four pieces of work nobody checked, and unchecked AI agent work usually costs more than work not done.

What raises that number

Three things, in order of effect:

• Visible state without opening. Working, waiting on you, finished, crashed. A glance costs a second; reading a terminal costs a minute.
• An alert when one stops. A stuck agent goes quiet, and quiet is indistinguishable from work. With no notification, you find out by accident.
• Tasks with a completion criterion. An agent that knows the command proving it is done fixes itself before calling you, and leaves your queue.

What lowers it

Overlapping tasks. Two AI agents in the same file overwrite each other with no error on screen, and the cost of finding out erases any parallelism gain. If the fronts are not independent, the number drops to one.

More agents or bigger tasks?

Depends how much review time you have. A bigger task produces a bigger delivery, which takes longer to check and is harder to reject when it comes out wrong. In practice three small independent fronts beat one large one, because you approve incrementally instead of choosing between accepting a huge block or redoing everything.

What if I just want to leave them running unattended?

Then the question shifts from "how many" to "with what guarantees". An unsupervised agent needs stronger isolation than a worktree (a container), an automatic completion criterion, and someone reviewing the result before it is merged. Without that, the right number is zero.