BrowserGym vs Real-Website Benchmarks

A great harness for research is not the same thing as a live-web readiness test

BrowserGym WorkArena Web Bench

BrowserGym shows up in almost every serious web-agent paper from the last two years. That is a good thing. It also creates a quiet category error: people treat “we evaluated in BrowserGym” as “we measured production browser automation.”

BrowserGym is a framework. Web Bench is a live-web task suite. Conflating them is like conflating Gymnasium with a specific robotics competition.

What BrowserGym is good at

ServiceNow Research built BrowserGym as an OpenAI Gym-style environment for multimodal web agents. You get:

  • Flexible observations (DOM / accessibility tree, screenshots, coordinates)
  • A broad action space, including high-level primitives and code
  • Chat-style interaction hooks
  • Compatibility with prior benchmarks such as WebArena and MiniWoB
  • Shared experiment tooling via AgentLab

That shared plumbing is why BrowserGym improved comparability. The WebArena paper’s GPT-4 setup and later BrowserGym-hosted runs are not the same experiment; the harness itself can change scores. Having one place to hold observation and action definitions constant is real scientific progress.

WorkArena, and the “Salesforce” search confusion

When search queries say “BrowserGym Salesforce benchmark,” they almost always mean WorkArena: knowledge-work tasks on a remote-hosted ServiceNow instance, run through BrowserGym.

WorkArena L1 focuses on atomic UI skills across the ServiceNow surface. WorkArena++ composes those skills into longer planning and memory problems. This is an excellent stress test for enterprise agents—dense forms, catalogs, filters, multi-step fulfillments.

It is still a controlled enterprise world. It is not 452 unrelated consumer domains with Cloudflare in front of half of them.

What the harness does not measure

Even when BrowserGym hosts WebArena-style sites or WorkArena’s ServiceNow stack, you typically do not measure:

  • Open-web bot detection and captcha loops
  • Residential proxy quality and ban rates
  • Marketing-site UI churn week to week
  • Write actions that can affect real third-party users

Those modes dominate real deployments. In our Web Bench study they showed up as a first-class error class; BrowserBench exists specifically to isolate the infrastructure slice (where we saw 25–50% accuracy swings from infrastructure alone).

When to report which number

Claim you want to make Prefer
Our planner / observation / memory change helped BrowserGym + WorkArena or WebArena
Our agent handles ServiceNow-style knowledge work WorkArena
Our agent works on the public web, including writes Web Bench
Our failures are stealth / proxy, not reasoning BrowserBench

A clean workflow we recommend to teams:

  1. Iterate fast in BrowserGym.
  2. Publish WorkArena / WebArena numbers for research comparability.
  3. Gate customer-facing live-web claims on Web Bench.
  4. If live scores crater, measure infrastructure before rewriting the agent.

BrowserGym made the field more rigorous. Web Bench is for the harder claim: that the agent still works when the website did not agree to be a gym.

More context: browser agent benchmarks map and WebArena vs Web Bench.