Browser Agent Reliability Lab

• [Playwright Trace Viewer](https://playwright.dev/docs/trace-viewer) already provides local DOM snapshots, network logs, console messages, and action timelines without uploading traces. It is a strong free substitute.
• [WebBench](https://webbench.ai/) and Browser Use benchmarks answer “which model or agent is stronger overall,” not “why did my production workflow fail three times out of ten?”
• [Browserbase](https://www.browserbase.com/pricing) and [Steel](https://steel.dev/) provide browser sessions, identities, observability, and production-scale infrastructure.
• Programmatic tools such as playwright-trace-analyzer show that parsing alone is not a moat.
• The open wedge is a multi-run acceptance gate: combine 3–20 runs of one business task, judge user-defined completion criteria, cluster failure steps, quantify variance/cost, and issue a release/hold decision.
• SERP conclusion: generic terms are occupied by official docs, vendors, comparison pages, and benchmarks. Exact Top 3/Top 10, monthly volume, and KD remain unmeasured. Do not attack AI browser automation as the primary generic keyword.

MVP Checklist

1. 1.Freeze the V1 bundle schema, privacy boundary, failure taxonomy, and 20 fixtures.
2. 2.Build pure TypeScript manifest/trace parsing and schema/version error handling.
3. 3.Build acceptance criteria, statistics, failure fingerprinting/clustering, and unit tests.
4. 4.Complete the sample Analyze-to-Results loop.
5. 5.Build the npm collector and real 2–20-bundle import.
6. 6.Add evidence drawer, release gate, next experiment, and Markdown/JSON export.
7. 7.Add `$79/$249` Payment Link/priced lead and commercial events.
8. 8.Complete Home, Methodology, Pricing, About, FAQ, and Privacy.
9. 9.Add mobile states, accessibility, errors, SEO metadata, FAQ schema, and Open Graph.
10. 10.Run build, Vitest, Playwright E2E, manual fixture review, and privacy checks.

Don't Build

• ✕Do not expand into a cloud browser or browser-automation platform.
• ✕Do not add complex backends, execution queues, proxy pools, credential systems, or CAPTCHA bypass.
• ✕Do not build custom auth, membership, orders, subscription billing, or admin first; keep lightweight payment validation.
• ✕Do not claim support for frameworks without fixtures and real trace validation.
• ✕Do not silently remove user-excluded runs.
• ✕Do not upload cookies, storage state, passwords, page bodies, or raw traces.
• ✕Do not replace reproducible evidence with LLM-generated prose.
• ✕Do not sacrifice launch speed for superficial completeness.