Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
The defensible way to benchmark a browser automation agent is to treat every benchmark as a separate experiment. Record the benchmark and revision, task set, environment, evaluator, model and agent versions, permissions, repeated-run results, cost, latency, and uncertainty. Publish raw per-task outcomes before any aggregate. Do not average percentages from unrelated leaderboards: as Steel’s methodology puts it, “a 92% on one benchmark and an 80% on another is not a ranking.”
What a browser-agent leaderboard should tell you
A leaderboard is useful only when its score has a precise meaning. A browser agent can succeed at a short, scripted interaction yet fail at a long workflow involving planning, authentication, changing pages, and recovery from errors. Your report should therefore answer five questions:
- What tasks? State the domains, task count, single-site or cross-site structure, and whether pages are synthetic, self-hosted, or live.
- What counts as success? Identify the official evaluator and whether it checks an exact answer, a final page state, partial credit, human judgment, or a hybrid.
- What agent ran? Name the model version, agent scaffold, prompts, browser, tools, permissions, and network conditions.
- How stable is the result? Give the number of attempts, uncertainty, failure categories, and run dates.
- What did it cost? Report runtime, model and tool calls, browser infrastructure, retries, and recovery work when those figures are available.
These details let a reader reproduce the experiment instead of treating one percentage as a universal measure of browser intelligence.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallWebArena, AssistantBench, BrowserGym and AgentLab: what differs
| Benchmark or framework | Environment and scope | What it is good for | Important qualification |
|---|---|---|---|
| WebArena | Self-hostable web environment for autonomous agents; the paper reports 812 tasks. | Controlled, multi-site web workflows with a reproducible environment. | The paper’s best GPT-4-based agent achieved 14.41% end-to-end task success versus 78.24% human performance. Those are historical 2023 paper results, not a current leaderboard claim. |
| AssistantBench | Live open-web evaluation with 214 tasks spanning more than 525 pages on 258 websites. | Long, realistic planning, navigation, and information-transfer workflows. | Live pages, logins, APIs, and anti-bot controls can change. Record the run date and environment state. |
| BrowserGym | Open, extensible framework that lists MiniWoB, WebArena, WebArenaVerified, VisualWebArena, WorkArena, AssistantBench, WebLINX, OpenApps, and TimeWarp. | Running several web-agent evaluations through a common interface. | A shared harness improves operations; it does not make scores interchangeable. |
| AgentLab | Tooling associated with BrowserGym for implementing agents, running evaluations, collecting traces, and analyzing results; its project description also mentions parallel BrowserGym experiments and unified leaderboard reporting. | Consistent experiment execution and trace analysis across benchmark adapters. | Unified reporting is a presentation aid, not evidence that task definitions or metrics are equivalent. |
Use WebArena when you need a self-hosted, controlled environment; use AssistantBench when open-web realism is the research question. BrowserGym and AgentLab are useful infrastructure around multiple task sets, not replacement benchmarks.
#1 Best Overall
Choose the benchmark before choosing the score
Separate environment classes
Put synthetic pages, self-hosted replicas, and live websites in different sections of your report. A self-hosted run can be repeated against a pinned snapshot; a live-web run is exposed to page redesigns, outages, changed search results, account state, and anti-bot defenses. Combining them into one percentage hides those differences.
Match interaction complexity to your claim
Single-step actions test precise interaction. Long-horizon tasks test planning, memory, navigation, information transfer, and recovery. Cross-site tasks add identity and state-management risks. State which category your conclusion covers; do not call a short-form score “general browser automation.”
Declare the evaluator
Prefer an official evaluator when one exists. Document the exact success predicate, partial-credit rules, timeouts, and treatment of blocked or ambiguous pages. If humans adjudicate outcomes, report the rubric and agreement process rather than presenting judgment as an exact measurement.
Recommended Free Tools
Build a defensible leaderboard, step by step
- Write a task-scope statement. Name the benchmark revision, task count, domains, page type, authentication requirements, and whether tasks are single-site or cross-site.
- Freeze the software context. Record model name and version, agent code revision, prompts, browser and driver versions, benchmark commit or release, enabled tools, permissions, proxy or network settings, and relevant environment variables.
- Define success before running. Store the evaluator version and the exact outcome categories: success, partial, failure, timeout, blocked, or infrastructure error. Never change the rule after seeing results.
- Run repeated trials. Use independent attempts where the benchmark permits it. Keep every per-task outcome and trace. Record seeds or sampling settings when applicable.
- Measure operational cost. Capture wall-clock latency, model tokens if available, browser/tool calls, infrastructure time, retries, and monetary cost. Explain whether failed attempts consume resources.
- Calculate uncertainty. Report the numerator and denominator, not only a rounded percentage. For a binary success rate, include a confidence interval or another stated uncertainty method, especially when the task set is small.
- Publish raw results first. Release a per-task table or machine-readable file, then show benchmark-specific aggregates. Remove credentials and personal data from traces.
- Track drift. Keep run dates, page or environment revisions, and a fixed audit subset. Rerun that subset after browser updates, benchmark changes, major site redesigns, or evaluator changes.
How to report scores without creating a false ranking
Keep each benchmark in its own column. A compact report can look like this:
| Benchmark | Agent/model | Tasks | Successful | Success rate | 95% interval or uncertainty | Median latency | Cost | Run date |
|---|---|---|---|---|---|---|---|---|
| WebArena (revision) | name and version | 812 or stated subset | raw count | count ÷ tasks | method and interval | measured value | defined accounting | UTC date |
| AssistantBench (revision) | name and version | 214 or stated subset | raw count | count ÷ tasks | method and interval | measured value | defined accounting | UTC date |
Do not put the WebArena percentage beside the AssistantBench percentage and label the larger one “better.” The tasks, environments, evaluators, and exposure to drift differ. If a combined index is required, publish the normalization formula, weighting rationale, missing-data treatment, and sensitivity analysis; retain the original scores beside it.
Useful secondary metrics
- Completion reliability: success over repeated attempts on the same task.
- Failure mix: navigation, perception, planning, tool-use, authentication, timeout, evaluator, and infrastructure failures.
- Efficiency: median and tail latency, browser actions, tool calls, and cost per successful task.
- Recovery: whether the agent detects a failed action and completes the task without human intervention.
A small, reproducible results calculator
The following Python program reads a CSV with an outcome column containing success or another value, then prints the rate and a Wilson 95% interval. It does not connect to a benchmark; it keeps aggregation separate from task execution so the evaluator’s output remains auditable.
#!/usr/bin/env python3
import csv, math, sys
if len(sys.argv) != 2:
raise SystemExit("usage: python summarize.py results.csv")
successes = 0
total = 0
with open(sys.argv[1], newline="", encoding="utf-8") as f:
for row in csv.DictReader(f):
total += 1
if row["outcome"].strip().lower() == "success":
successes += 1
if total == 0:
raise SystemExit("no rows found")
p = successes / total
z = 1.96
center = (p + z*z/(2*total)) / (1 + z*z/total)
half = z * math.sqrt((p*(1-p) + z*z/(4*total)) / total) / (1 + z*z/total)
print(f"tasks={total}")
print(f"successes={successes}")
print(f"success_rate={p:.4%}")
print(f"wilson_95={max(0, center-half):.4%}..{min(1, center+half):.4%}")Keep the CSV row for every task, including failures and infrastructure errors, and define in your report whether those errors remain in the denominator. If you exclude them, publish both the attempted-task rate and the conditional rate.
Cost, latency and reliability details readers often miss
Cost accounting
State whether your cost includes only model usage or also browser hosting, proxy traffic, storage, retries, and human intervention. A cheap successful run can be less useful than a slightly more expensive run if it requires frequent manual recovery.
Latency accounting
Report at least a median and a high percentile when possible. Clarify whether time spent waiting for page loads, evaluator calls, retries, and queueing is included. One unusually fast cached page should not represent an entire live-web workflow.
Reliability accounting
Separate agent failures from environment failures. A CAPTCHA, unavailable API, expired login, blank page, or benchmark-service outage should have its own category. Otherwise a leaderboard may punish an agent for an event it could not control—or hide a brittle agent behind “infrastructure error.”
Environment drift and audit practice
Live evaluations need a maintenance policy. Record the UTC date, browser build, benchmark revision, page URLs or task identifiers, account state, and network region used for every run. Maintain a small fixed audit subset and rerun it after site redesigns, benchmark updates, browser upgrades, evaluator changes, or new anti-bot behavior. For self-hosted environments, pin images, databases, seeded data, and configuration so another researcher can recreate the same state.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →When a task becomes impossible because a page or API disappeared, do not silently delete it. Mark it unavailable, explain the disposition, and show how the denominator changed. A leaderboard with fewer tasks after an update is not directly comparable to its earlier version.
Common benchmarking failures and fixes
- “The highest percentage wins.” Fix: compare only within the same benchmark revision and evaluator; show separate benchmark tables.
- Scores are rounded with no counts. Fix: publish successes and total attempts so readers can reconstruct the rate.
- One run is treated as definitive. Fix: repeat trials, preserve traces, and report uncertainty and failure variance.
- Tool access is omitted. Fix: list browser actions, search, code execution, file access, credentials, and network permissions for every agent.
- Benchmark and agent versions are missing. Fix: pin commits or releases and include prompts and configuration in an experiment manifest.
- Live-site outages are counted as agent mistakes. Fix: classify environment failures separately and publish the rule before running.
- Human and automated results use different evaluators. Fix: state the human protocol and do not compare unlike measurements as if they were identical.
- Traces leak secrets. Fix: redact cookies, tokens, personal data, and private page content before release.
Or skip the browser setup
For documenting benchmark pages, ScreenshotNeo is the first alternative to try: it removes consent banners, newsletter popups, and chat widgets before capture, and bills only clean shots.
One GET request returns a PNG, JPEG, WebP, or PDF. See the complete parameter list in the ScreenshotNeo documentation.
curl -G "https://api.screenshotneo.com/v1/shot" -d access_key=YOUR_API_KEY --data-urlencode url=https://stripe.com -o shot.webpimport requests
r = requests.get("https://api.screenshotneo.com/v1/shot", params={"access_key": "YOUR_API_KEY", "url": "https://stripe.com"}, timeout=90)
open("shot.webp", "wb").write(r.content)const q = new URLSearchParams({ access_key: 'YOUR_API_KEY', url: 'https://stripe.com' });
const res = await fetch(`https://api.screenshotneo.com/v1/shot?${q}`);For benchmark evidence, useful options include full-page capture with lazy images loaded, a CSS-selected element, dark mode, device presets or a custom viewport, retina scale, PDF paper size and page ranges, custom CSS or JavaScript, clicking before capture, hiding selectors, waits for a selector, delay, or network idle, blocking ads, trackers, requests, or resource types, custom headers, cookies, user agent and Authorization, timezone and geolocation, transparent backgrounds, resizing, chosen cache TTL, signed links, asynchronous jobs with signed webhooks, bulk capture of up to 100 URLs per call, a usage API, and an OpenAPI specification. Parameter names used by other screenshot APIs also work, which can simplify migration.
Bot checks or CAPTCHAs, blank pages, timeouts, failed loads, and cache hits are not billed; response headers identify the page verdict and billing status. ScreenshotNeo also provides an MCP server with take_screenshot, get_page_info, and capture_pdf for Claude, Cursor, and other MCP clients.
| Plan | Price | Included shots |
|---|---|---|
| Free | $0 | 1,000 per month, no card |
| Starter | $5 | 3,000 |
| Growth | $15 | 15,000 |
| Pro | $39 | 60,000 |
| Scale | $99 | 250,000 |
| Business | $249 | 1,000,000 |
Every feature is available on every plan, and yearly billing gives two months free. You can start with 1,000 free screenshots a month with no card; paid plans start at $5 for 3,000 shots.
Frequently Asked Questions
How often should a browser-agent leaderboard be rerun?
Rerun a fixed audit subset after any benchmark, evaluator, browser, model, site, login, or anti-bot change, and retain the original run rather than overwriting it.
Can I publish traces from tasks that contain private data?
Only after removing credentials, cookies, tokens, personal information, and restricted page content; if redaction would make the trace misleading, publish aggregate outcomes and the redaction policy instead.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What should I do when a benchmark task disappears?
Mark it unavailable, record the reason and date, preserve the original denominator, and show any revised denominator separately so historical scores remain interpretable.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
