← Monaco 2022

Ep 05Monaco 2022

What if the pit wall let the AI act?

The same eight laps as episode 3, but Rosso works for you. It watches, changes the view and makes calls. You decide how much it's allowed to do.

F1AI agentsInteraction design

Episode 3 was the pit cycle as it happened: Carlos leading on wets, the double stack, Latifi. This episode replays the exact same eight laps, but you're on the Ferrari pit wall and Rosso works for you. Not as a chatbot you ask questions. As an agent that watches, decides and acts.

The race never changes. The cars do exactly what they did on the day. Rosso controls the view and makes calls, and every call is checked against what really happened.

Built for a laptop. Strategists don't make race calls from their phones. Open full screen ↗

How much is it allowed to do?

That's the real design question with any AI agent, so it's a dial:

  • Suggest: Rosso never touches anything. It only tells you what it would do.
  • Ask first: it changes the view by itself, with an undo, but pauses the race and waits for you before any race call. Approve with A, reject with R.
  • Act on its own: it does both, with a few seconds to undo.

Risk decides what it asks about. Moving the camera is cheap, so it just does it. A race call isn't, so it asks. And your settings always win: pick a camera yourself and Rosso stops moving it. Press Stop Rosso and it does nothing else, but its activity trail keeps everything it already did.

Its calls, checked against the day

  • Keep Carlos out, skip inters, go straight to slicks. Same as the day. Medium confidence.
  • Don't double stack. Different from the day. Medium confidence. Red Bull double stacked too, a lap later.
  • Ask race control to show Latifi blue flags. Same as the day. High confidence.
  • Prediction: Pérez rejoins ahead of Carlos. Right, by 2.2 seconds. High confidence.

Whose confidence?

At first the labels just said high, medium or low, and I had to stop and ask: whose confidence? Mine, or the agent's? So now every label says Rosso's confidence, and the human never rates anything. You approve or reject.

Rosso follows a fixed rule, written once at the bottom of the page so you never have to read it twice. High means the call follows directly from numbers already measured. Medium means it depends on something that can't be measured yet. Low means an educated guess.

One screen, no scrolling

My first version stacked everything down the page. Then it hit me: a strategist can't scroll up and down to find the cars. By the time you've scrolled, they're somewhere else. So the tower, the map and Rosso's panel sit side by side on one screen, Rosso's activity scrolls inside its own panel, and Approve and Reject sit right on the map, where your eyes already are.

I also took out the function calls. My early agent showed lines like pitwall.call(simulated). A strategist doesn't care which function an AI called. They care what it noticed and what it decided.

Personal concept. Not affiliated with Formula 1 or Scuderia Ferrari.