AgentGauntlet — crash-test your browser agent

Agent

What is being tested. AgentGauntlet doesn't care how it thinks — only whether it survives.

Reference AgentDeterministic built-in agent. No LLM API key required. Handles: dismissOverlays, waitForLateElements.Naive AgentDeterministic. No overlay handling, no waiting — the control arm.Resilient AgentDeterministic built-in agent. No LLM API key required. Handles: dismissOverlays, waitForLateElements, recoverSessions.Example Repository AgentRun your own agent inside an isolated Solari Sandbox.LLM AgentOptional adapter. Needs a model API key; nothing else here does.

Task

What the agent has to accomplish, and how it will be judged.

Task

Add Aurora Headphones to cart, apply coupon SAVE20, proceed to checkout, enter Name: Ada Lovelace and City: London, continue to review, and stop before submitting payment.

max 25 steps · 90s timeout

Environment variants

Each one changes the environment deterministically, seeded from the run's coordinates.

Repetitions

The same variant, run this many times. Repetition is where unreliability shows up.

2