What Is an AI Agent Controller?

An AI agent controller is not a category anyone has formally defined, so here is the plain version: it is the small hardware you keep within reach because you now spend your day approving, rejecting and re-prompting an AI agent. Slam Buttons, two large arcade buttons on a cream plate for $29, is the whole idea in its crudest form. Bind one to accept and one to reject, and you stop hunting for a key you press hundreds of times a day.

Why has this become a thing to buy?

Because the bottleneck moved. When the model writes the code, you become the reviewer, and reviewing is a small set of actions repeated at a rate no one designed a keyboard for. Accept. Reject. Interrupt. Scroll. Ask again. None of those are hard, but doing them from scattered keys, with a hand leaving the home row each time, is the part of the day that quietly wears.

What are the parts?

Four, and most desks want one or two of them, not the set.

• Something to hit. Slam Buttons at $29, or the Pedal Three at $69 if you would rather your feet did it. Both send whatever key you assign.
• Something to type shortcuts. The KM16 Macropad at $119 for sixteen keys and three knobs, or the EM Mini Pad at $75 for five keys and a knob when desk space is the constraint.
• Something to turn. The knob on either pad, which is the part most owners end up using most: agent output is long and scrolling is continuous.
• Something to look at. The Status Light at $45 or the Desk Terminal at $79, so you can stop switching windows to find out whether the run is done.

Does an agent controller integrate with the agent?

No, and the first three parts do not need to. They are keyboards in a different shape: they send keystrokes, the agent reads keystrokes, nothing negotiates. The light is the one exception, because a light has to be told something. It is powered over USB-C and set by a hook or a script you write at the start and end of a run, which is easy with a terminal agent and more work inside an editor. If you want a device that reads an agent's state by itself, nothing here does that.

Do I need a different controller for each tool?

No, and this is the main argument for buying generic hardware. The same pedal, pad, knob and light work with Claude Code, Codex, Cursor, Copilot, Gemini CLI and Windsurf, because all any of those see is an ordinary key press. A controller built for one vendor can do two things this cannot: carry that tool's own artwork, and be handed the tool's status directly. It also stops being useful the day your work moves to another tool, which in this field is roughly every few months.

What will it not fix?

It will not make you a better reviewer. Making approval the easiest, most satisfying action on the desk is not obviously good for the quality of what you approve, and that is worth sitting with before you buy the biggest button. It will not reduce the number of decisions either. And if you share a wall or a room, arcade switches and mechanical keys are loud, which is a real cost that none of the quiet options ask you to pay.

If the accept and reject half is where you want to start, read our note on a physical button for approving what your agent just did.