Same task. Different systems. Measured.

Run one task through different agent setups and see which is most reliable — and what it costs. Click any system to change its model, instructions, tools, or how its agents coordinate.

History
Your task
What does a correct answer look like? optional

Each system runs your task several times. A judge model grades every answer — reliability is how often it passes.

Systems under test
tap a system to include or exclude it
Single agentsingle agentGemini 3 Flash

One model, one pass. No second agent checks the answer.

Configure →
Sequential reviewteam2 agents

Your prompt goes to the first agent; its answer is passed to the next to check and improve.

Configure →

Each system runs the task several times. Scored on reliability (how often it's right) and cost.

Experiment tracks · the summer program

Three kinds of experiment a team can take on. Pick one, run it here, share what you find.

01runnable now
Known-answer

Problems with a ground truth, so any team can verify the result independently.

  • · One typo. Who catches it?
  • · The attack: an instruction hidden in the document
02open to every team
Bring your own problem

Put your own task or agent system under test. A judge model grades each answer against what you define as correct.

  • · Use “test your own task” above
03forming
Public-impact

A high-visibility experiment on a real public problem, like wildfire or coastal-safety risk, with regional partners.

  • · In design — bring a problem set