You're paying two thousand dollars a month for the most capable machine ever built, and the clever new trick is teaching it to talk like a caveman.
I'm not exaggerating. Go look. There are repos littering GitHub right now that grind your agents' language down to grunts to cut the token bill. Caveman SpongeBob. Kevin from the office. The whole pitch is that agents waste money talking to each other, so the move is to make them talk worse. Pay frontier prices, receive cave paintings.
And here's the part that gets me. The instinct underneath it is correct.
Agent chatter is expensive. When you wire up a crowd of models to coordinate, most of what they spend gets spent on talking about the work instead of doing it. Cutting that down is real engineering, and anyone who's ever read a real bill knows the back-and-forth is where the money goes. So I'm not going to stand here and tell you tokens don't matter. They matter. We count every one of them.
What I'm telling you is that grunting isn't compression. It's degradation.
A caveman sentence is shorter and dumber, and dumber has a price. The agent on the other end can't parse mE rEdO tHe PaCkEtS, so it guesses, or it asks again, and now you've paid for the short message and the round trip you were trying to avoid. You didn't save anything. You bought ambiguity at a discount and got billed for it twice. You took the smartest tool you will ever own and asked it to be stupid on purpose. On purpose.
There's a better answer, and it's been hiding in a movie everyone already loves.
Ratatouille. A Pixar cartoon, sure, but look past the rat. Remy can't run a kitchen, so he puppets Linguini, the lowliest cook on the line, and through him you watch a real French brigade run a full service. Behind the gag is the one organization on earth that solved this exact problem a hundred years before we had it. A professional kitchen on a Friday night is a dozen specialists, a brutal clock, zero tolerance for a dropped plate, and a bill that punishes every wasted motion. Sound like anything you're building? They make a crowd move like one body, and they do it on language.
My lovely wife and cofounder, Stacy Silva, wrote the other half of this. She spent fifteen years on professional lines before she built software with me, and she'll tell you the brigade is the architecture: one expediter, every cook on one station, nothing leaving the kitchen until it crosses the pass. Go readher piece. This is the same kitchen, one shift later. She gave you the structure. I'm here for the language.
Because the kitchen didn't get terse by getting dumb. It got terse by getting professional. Stand next to a real line under fire and you'll hear maybe six words all night, and every one of them is loaded.
"Heard." I got it, I understood it, it's mine now. Nothing moves on faith. Every call gets a word back.
"Behind." "Corner." I'm at your back with a hot pan, don't turn. Position and danger in one syllable.
"Fire." Start it now, on my timing.
"86." It's gone. Kill it. Stop.
"Hands." Stop the line. This one is urgent and everybody moves on it now.
"Walking in." Heard, I'll fold the new order into the flow without dropping what's already on.
"In the weeds." I'm underwater and the system needs to rebalance before I take you down with me.
That's not caveman. That's the opposite of caveman. Every word carries state, priority, and an action, with no ambiguity left over to pay for later. It is the most compressed professional language I know, and it's compressed because a hundred years of people who literally could not afford to be misunderstood sanded every wasted syllable off of it. You're not going to out-engineer that with a meme filter. The work is done. They wore checked pants while they did it.
So when Stacy and I built our system, we didn't invent a clever little compression scheme. We robbed the kitchen, on purpose, top to bottom.
We don't run pods. We retired the word. We run stations, and each one is a specialist that owns its job and never reaches into anyone else's. Work moves between them in handoffs that get checked on both ends, so nothing advances on a message nobody confirmed. That's "Heard," in code. Our agents run a nine-step pipeline and exactly one of those steps is allowed to talk to the model. The other eight are deterministic. The expensive, unpredictable part is boxed into a single station, the way a kitchen boxes the fire to the people standing at the range.
We meter every token a step is about to spend before it spends it. When a run leans on its budget, the line tightens up and compresses what it's carrying. It does not start grunting. When a run tries to blow its budget, we 86 it. The run gets held and the reason gets logged in plain language: budget exhausted. Work crosses a pass, gets inspected by critics that are only allowed to say pass, fail, or escalate, and it either gets sealed or sent back. The pass fails closed. There is no override on the seal. The standard is the product.
We named the whole thing the Brigade, and we taught it to talk like one. Heard. Hands. Fire. 86. Walking in. Real words, from a real craft, because the words already existed and they already won.
Here's the part I want you to sit with. Our entire field has a habit. We discover something hard, find out humans solved it a century ago, and then refuse to go learn how, because taking notes from a line cook feels beneath a person who owns a frontier model. So instead of studying the most battle-tested coordination system on the planet, we made a cartoon out of it. We had the answer and we turned it into a meme.
The people who run kitchens have done low-latency, high-stakes, fault-tolerant coordination, under a clock, on a budget, with a crew that turns over every month, since before your favorite framework's grandparents compiled. They are not cute. They are the reference implementation.
You can keep paying premium prices to talk to a caveman. Your call. We taught ours to talk like "Cousin."
Heard.