Benchmark Controls
Run the benchmark to compare task success, average step latency, tool-call success, and intervention count.
Workspace Snapshot
Profile Matrix
Winner Draft
No benchmark run yet.
AI WebGPU Lab Benchmark
Agent Step Latency Benchmark
`bench-agent-step-latency` compares deterministic browser agent profiles against one fixed local task deck, tool catalog, and scoring rule.
The benchmark keeps task completion, tool-call success, step latency, and intervention count in one schema-aligned result surface before real browser-controller runtime wiring lands.
Task Deck
Metrics
Environment
Fixture Metadata
Loading fixture...
Activity Log
Schema-Aligned Result Draft
{
"status": "pending"
}
Benchmark Notes
- All profiles share one task deck, local-only workspace, and tool catalog.
- The winner favors high task success and tool success before lower step latency.
- Real browser controller and planner runtimes should keep the same metric fields and artifact names.