Contents
TL;DR
Startup RankCaster AI replaced manual regression with an autonomous AI agent: deployed in one day without Selenium/Cypress, cut runs from 48 hours to 10–40 minutes, and cost per full pass from ~$250 to ~$5.
What happened
At Pre-Seed scale, every release inflated COGS: more features meant more QA scenarios. The team stopped paying for dashboard clicking and built a three-part stack:
| Component | Role |
|---|---|
| Claude Code | Decisions from visual context |
| agent-browser (Rust + Chrome CDP) | Controls real Chrome via CLI |
| SSH / psql (read-only) | Validates UI against test DB |
The agent reads the Accessibility Tree, not pixel coordinates—changing a button ID or color does not break flows. After clicking a filter it runs SQL and compares chart numbers to table data.
In the first week it found bugs humans missed: an APR calculation error under rare filter combinations (API vs DB mismatch) and CSS clipping on charts.
Why it matters
This is not “replace the QA team”—it shifts the model: routine regression and first-pass root-cause analysis move to automation; engineers lock findings into deterministic Vitest tests.
The approach skips classic E2E frameworks: semantic navigation plus direct DB checks in one pass across UI, API, and data.
In practice
- Isolated QA accounts on a beta environment and one “memory” file with command patterns—no per-button scripts.
- Data masking: the agent uses a shadow DB copy; PII is replaced by ETL before copying from staging.
- In GitHub Actions each run is an isolated container; data mismatches block release.
- Canvas and heavy visualization remain weak spots; the agent covers critical paths and logical anomalies, humans handle fine UI polish.
- Every found bug becomes a Vitest test so you do not pay twice for the same scenario.
Bottom line
The Habr case is a practical agentic QA example in 2026: cheaper than manual regression, faster reports (screenshot, log, SQL). Full replacement of testers is not claimed—but release economics change noticeably.