Research
We build the environments and benchmarks that teach AI agents to do real financial work—the documents, models, and multi-step judgment calls behind the most economically valuable tasks—and we publish what we learn about where frontier models succeed and where they break.
Research areas: Agent Benchmarks · Long-horizon Finance Evaluation · RL Environments
DealTrace Bench
A long-horizon review of a real private-equity deal in five graded stages, tracing where models are strong and where they break. Failures concentrate in forecasting—and forecast quality most limits end-to-end performance.
Read the post →Agent Benchmarks
What can browser and computer-use agents actually do on the open web?
BrowserBench
How bot-detection systems distort browser-agent evaluation, with stealth vs. non-stealth performance across real infrastructure.
Web Bench
A benchmark of ~2,454 realistic tasks across 452 real-world websites for evaluating browser agents.
Long-horizon Finance Evaluation
Where do models break inside multi-stage analytical work?
DealTrace Bench
A staged review of a real private-equity deal in five graded stages—extract, reconcile, forecast, market, recommend.
RL Environments
Where can agents practice consequential work without real-world side effects?
Westworld v1
AI sandbox environments for training and evaluating computer-use agents on realistic simulated websites.