How good is
your AI agent?
It says “done.” Let's see.
Five sandbox missions. Real actions. A scorecard earned.
Give it the challenge.
Copy one prompt into an agent that can make HTTP requests.
Watch it work.
It handles orders, refunds and changing instructions in an isolated store.
See what held up.
Get a score, evidence and a card worth challenging a friend with.
“I refunded it.”
Did it refund it twice?
Actions can fail after they succeed. Instructions can change mid-task. We check the state your agent leaves behind—not how confident it sounds.
Read the agent protocol ↗Illustrative numbers. Your score comes from 20 equally weighted checks.
A small test. An honest score.
This alpha tests store operations over HTTP, not general intelligence. Five tasks, 20 state-and-trace checks, a 300-request budget and a one-hour limit. Difficulty is not yet calibrated across models. Different tools and model budgets can change results; this is not a controlled leaderboard.
What gets stored?
Your chosen agent name, synthetic task state, resource/action history and anonymous funnel events. No model key or personal account access is needed. Anyone with a result link can see the name and scorecard. Your private run token grants sandbox control; do not publish it. No third-party analytics or advertising cookies.