Resurf

Summary

Testing browser agents today means choosing between flaky real websites and static HTML benchmarks missing state and dynamics. Resurf offers a middle path: a synthetic e-commerce site that replays deterministically and injects failure modes.

Resurf is a framework for reproducible testing of AI browser agents built on a single synthetic site (shop_v1) that exercises forms, multi-step checkout, auth, and inventory logic. The tool records per-step DOM snapshots, screenshots, and token counts. It supports multiple agent adapters (browser-use, stagehand) and runs modifiers—failure-mode injectors for network latency, payment declines, and server errors—via YAML configuration without code changes. The core appeal is reproducibility: tests replay from SQLite snapshots with seeded randomness rather than hitting live sites. The limitations are structural: v0 ships one site only, the project has minimal GitHub activity (5 stars, 31 commits), and no independent benchmarks or community validation exist yet. This is genuinely early stage—suited only for teams willing to build on foundational abstractions before the ecosystem exists.

Bottom line: Pick this if your team needs deterministic agent testing and can accept a single e-commerce template. Once you need multiple domains or production-scale evaluation, you'll outgrow it quickly.