Your brain doesn't gate what comes in. It gates what gets out.

A working-memory habit we almost never question: we treat holding a thing in mind as the hard part. Get it in there, and surely it’s available to use. A study out this September pokes a hole in that. Natalie Nielsen, Roshan Cools, and colleagues at the Donders Institute gave people a working-memory task and briefly blocked a dopamine receptor (with sulpiride, a D2 antagonist). The information still went in fine. What broke was the brain’s ability to pick the right held item and push it out into action — and when that output step faltered, the representation in visual cortex got blurrier and task accuracy dropped with it. (One pharmacology study; treat it as a strong data point, not a verdict.)

That lines up with a framework cognitive scientists have been building for years: working memory has two gates, not one. An input gate decides what gets to enter and stick. An output gate decides which of the things you’re already holding gets to drive what you do next. They’re separable. You can be holding exactly the right fact and still fail because the output gate opened on the wrong one.

Sit with that, because it quietly inverts how most of us build.

We pour our effort into the input gate. What do we put in the context window. What do we show on the screen. What do we retrieve. The whole conversation is about getting the right stuff in front of the system — and then we assume the rest takes care of itself, that if the relevant thing is present, it’ll be used.

The brain is telling us the assumption is wrong. Presence isn’t use. The scarcer, higher-leverage control is the output gate: of everything currently held, which single thing is allowed to act right now.

This is the whole ballgame for agents. An agent can hold an enormous amount — a giant context, a pile of retrieved documents, a dozen tools. The failures rarely come from not having the fact. They come from acting on the wrong held thing at the wrong step: calling the tool that was available instead of the one the moment needed, citing the chunk that was retrieved instead of the one that answered the question. That’s an output-gating failure, and you don’t fix it by stuffing more into the input gate. You usually make it worse.

So build the output gate on purpose. A few moves follow from this:

Separate “available” from “selected” in your own design. If retrieval and use are the same step, you have no output gate — the first plausible match acts. Put an explicit selection step between here’s what we’re holding and here’s what we’ll act on.

Make the gate inspectable. The reason output-gating failures are maddening is that they’re invisible: everything needed was right there. Log not just what the system retrieved or held, but which item it chose to act on, and why. You can’t debug a gate you can’t see.

Spend the expensive judgment at the output, not the input. It’s tempting to put your smartest filtering up front. But the brain reserves a neuromodulator for the output decision — the moment of acting is where precision seems to matter most. Mirror that: a cheap, generous input gate feeding a careful output gate often beats a paranoid input gate feeding a careless one.

The headline version of AI progress is about bigger context — hold more, see more, remember more. Useful. But capacity was never the thing that made a brain effective. A brain that held everything and gated nothing would be useless. The competence lives in the gate on the way out — the half-second decision about which of the many things you’re holding gets to move your hands. Build that gate, and build it where you can watch it work.