QA code review with AI agents

PR review is the phase where QA teams most often collapse — either it does not run properly (deadline pressure), or it steals so much senior time that they do not get to writing tests themselves. An AI review agent is a compromise: it does not replace human review but catches 50–70% of issues before they reach a person.

How to set up a QA review agent in Claude Code

Create the file .claude/agents/qa-reviewer.md:

---
name: qa-reviewer
description: PR review pre QA testy. Kontroluje Page Object
  patterns, anti-flaky best practices, pokrytie negative scenárov
  a duplicity. Spúšťaj pri každom PR s test súbormi.
---

Si QA reviewer senior level. Prechádzaš diff pridaných/
zmenených test súborov a kontroluješ:

## Kritické (blokuj merge)
- Hardcoded timeouts (cy.wait(number))
- Missing data-qa selektory — nikdy CSS class alebo ID
- Cez-test dependencies (state leak z predchádzajúceho testu)
- Missing beforeEach/afterEach cleanup
- Tested len happy path, bez negative scenárov tam, kde dáva zmysel

## Warning (nezjednostavuje)
- Duplikát existujúceho test case-u (prezrel si cypress/e2e/)
- Inline test data namiesto factory
- Nesprávny súborový zaraďovací (feature patrí do cypress/e2e/X, nie do Y)

## Nitpick (len mention, nezjednostavuje)
- Neosadené aria-labely v page objecte
- Chýbajúci JSDoc pri exportovaných helperoch

Výstup formátuj ako GitHub review:
- **file.ts:23** [CRITICAL] hardcoded timeout, použij cy.intercept()
- **file.ts:45** [WARN] duplikát cypress/e2e/checkout/payment.spec.ts:12

Running the agent

On a PR run in the terminal:

git checkout feature/new-checkout-tests
claude /agents qa-reviewer

> Reviewuj diff oproti main pre všetky .spec.ts súbory.

Claude reads the diff, applies the rules from the file and prints a review. Paste it into the GitHub PR.

CI integration (advanced)

Automatic comment on a PR via GitHub Action:

name: QA Review
on:
  pull_request:
    paths: ['cypress/**/*.spec.ts']

jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with: { fetch-depth: 0 }

      - name: Run Claude review
        uses: anthropics/claude-code-action@v1
        with:
          api-key: ${{ secrets.ANTHROPIC_API_KEY }}
          agent: qa-reviewer
          diff: ${{ github.event.pull_request.base.sha }}..HEAD

      - name: Post comment
        uses: peter-evans/create-or-update-comment@v4
        with:
          issue-number: ${{ github.event.pull_request.number }}
          body-path: .claude/review-output.md

Measurable delivery

• Human reviewer bandwidth — senior QA doesn't need to look at every trivial PR. They only look at PRs where the AI agent said 'LGTM but check one thing'.
• Framework stability — anti-flaky patterns are followed consistently, not only when a reviewer happens to notice.
• Junior onboarding — a junior tester gets immediate feedback on every PR. They learn patterns faster than from manuals.

What the AI reviewer won't solve

• Business logic — 'are we testing the right scenario?' is a human job.
• Architectural decisions — 'does this belong in a page object or is it a shared helper?' AI recommends but not authoritatively.
• Trade-offs — test speed vs. thoroughness is a business decision.

Best practice: 2-step review

1. AI review agent — automatically on the PR. Catches technical details.
2. Human review — a senior QA or dev lead takes at most 5 min. Focuses only on business and architectural correctness.

Real time needed to review a PR drops from 25–45 min to 5–10 min, with higher quality.

---

Want the same approach at your company? Get in touch — dohodneme 30-minute discovery call.