Aeroptimize

Multi-agent wing optimization: replay

Recorded run

Run

Start design

drag a dial

Live run

local server

Agents

Model

Access code

Evaluations

Cost cap ($)

Hybrid: Optuna inner loop allowed

Target

Run

GEN0/ 0

Cl—

Cd—

Solver—

Stall margin—

Cl–Cd polar

Agent decisions

verbatim from traces.jsonl

Aeroptimize · how to read this replay

A team of AI agents designs a race-car wing

Three language-model agents take turns improving a wing section until it makes a target amount of downforce without stalling. Every number comes from physics solvers and automatic checks, not from the agents.

The goal

Chiefreads the results so far and picks the next move
CADchanges the shape: camber, thickness, angle
Solversmeasure it: NeuralFoil (fast) and XFoil (accurate)
Criticchecks the result and diagnoses what is wrong

That loop is one generation. It repeats until a design passes or the budget runs out.

Wing (centre): the current design, the previous one dashed, and its downforce against the target.

Polar (below): every design so far. The aim is the green box.

Feed (right): each agent's reasoning in its own words.

Timeline (bottom): play, pause, step or scrub. Hover any underlined term for an explanation.