Everyone loves a breakaway.
AI made the attack free. It didn't make the agent finish the race.
§ 01The rider with the camera
The most exciting thing in a bike race is the breakaway. One rider jumps off the front. She gets the camera. She gets the commentary. She gets the crowd. For a while she looks like she's winning.
She almost never wins.
The race that counts is the general classification. Add up every stage, and the winner is whoever has the lowest total time at the end. Not the rider who did the most dramatic thing on a given afternoon. The rider who was quietly, relentlessly, boringly good for the whole race, with a team grinding in front of her so she could spend less energy doing it.
Demi Vollering wins bike races this way. She won the Tour de France Femmes in 2023 by being the best over the whole race. She won the Vuelta Femenina in 2024 and again in 2025 the same way. The highlight reel is short. The work is long.
She also lost the 2024 Tour by four seconds. A week of racing, decided by four seconds. That's what the general classification is. Nobody remembers four seconds as a highlight. Every one of them was won or lost somewhere nobody was filming.
Then 2026. In June she won the Giro d'Italia Women, which made her a winner of all three women's Grand Tours. In August she won the Tour again. Two years after losing it by four seconds, she was back at the start of the same race, doing the same long, unglamorous thing. That's the other part of the GC. Nobody gets to be done.
§ 02I am a breakaway specialist
I'm telling you this because I'm a product manager, and product managers love a breakaway.
We love the attack. We love the demo that gets the room. We love the launch with the countdown and the Slack channel and the screenshot in the all-hands. I have personally jumped off the front, gotten the camera, and called it a product. My roadmaps have been, if I'm honest, a list of breakaways. Each one thrilling. Most of them caught by the peloton within the quarter.
§ 03The attack got cheap
AI made this so much worse.
In "Show, don't tell" I wrote that AI collapsed the cost of showing. I still think that's the best thing to happen to the job. This is the bill.
A year ago, showing took effort. You had to build enough of the thing to show it. Now you can have an agent doing the impressive thing by lunch. The cost of the attack went to zero. So naturally we attack constantly. Five breakaways before standup. Look, it books the meeting. Look, it answers the ticket. Look, it writes the thing. Everyone claps.
The demo is real. The demo is also, almost always, a breakaway, and the peloton is right there.
When everyone can attack, an attack stops meaning much. A demo used to be evidence that a team could build the thing. Now it's evidence that a team had an afternoon.
§ 04The part that didn't get cheaper
The gap between an agent that does the thing once, in front of you, on the happy path, and an agent that does the thing every time, for everyone, on the paths you didn't think of, is enormous. The model doesn't close that gap. The model is the strong rider. The gap gets closed by the grind nobody films. The evals. The data you keep clean. The hundred boring failures you fix one at a time.
I've made pieces of this argument before. A demo is a single run you got to choose ("Your agent works. Prove it."). The agent logic is usually fine and the information underneath it isn't ("Nobody gave it the address."). This is the same point wearing lycra. The grind is the GC. It's long. It doesn't photograph. It wins the race.
The hard part is that the breakaway feels like progress and the GC feels like nothing. Cleaning the data isn't a launch. Fixing the fortieth edge case doesn't book a meeting. You do it, and at the end of the day the demo looks exactly like it did this morning. Which is to say it works. Which is the entire point. Which nobody will clap for.
§ 05Ride for the GC
So this is the job now. Resist the attack, or at least stop counting attacks as results. Do the boring accumulation. Let the agent be the protected rider, and make it actually win, not just look good on the one climb where the camera happens to be.
In fairness, Vollering won this year's Giro with an attack. She started the last stage behind Anna van der Breggen, who was wearing the leader's pink jersey. On the first climb, her team split the race and left Van der Breggen without teammates. One of them, Lauren Dickson, kept the leaders within reach. Near the top of the last climb, Vollering attacked, rode across to the front alone, and took the Giro by a minute and 37 seconds. She didn't even win the stage.
So yes, the race was won by an attack. Look at what it sat on. Every stage before it, she kept herself in the race. Her team spent the whole last day making the move possible. The attack wasn't the plan. It was the last step of the plan.
That's the only kind of demo I'm for now. The one that comes at the end, on top of the boring part, when the work has already made it safe to go.
Cycling has a name for the people who do that work: domestiques. They fetch bottles, block the wind, chase down other people's attacks, and often finish well back. Their names aren't on the trophy. The leader doesn't win without them. If the agent is the leader now, I know which job I have.
I'm aware, by the way, that I've just written an essay built entirely on a cycling metaphor, which is itself a breakaway. Flashy. Off the front. Very pleased with itself. There's a real chance the peloton catches it before the final paragraph.
But you read this far. So maybe it stays away to the line.
Probably not.