QAgentic

Manual regression testing couldn't keep up with the release calendar. An e-commerce brand was shipping frequent site updates, but every release meant hours of manual click-through testing across product pages, cart flows and checkout, across multiple locales and currencies. A single QA engineer couldn't cover every path before launch, so bugs slipped through: broken links, wrong pricing at checkout, mistranslated UI copy.

Scripted test suites had been tried before and abandoned. Selector-based tests broke every time the site's markup shifted, and maintaining them ate more time than it saved. The team needed tests that survived a redesign, not ones that broke with it.

Agents that test intent, not selectors. QAgentic runs tests written as plain-language flows ("add a medium black hoodie to cart and check out") and executes them against the live site the way a real shopper would. I built the system end to end: the agent execution engine, the test-plan builder, the evidence and reporting layer, and the CI integration that gates deploys on test results.

Instead of relying on brittle CSS selectors, the agent reads the page the way a person does and adapts to pop-ups, out-of-stock variants and promo banners on the fly. Every run produces a full step-by-step trace with screenshots, so a failure is immediately explainable rather than a mystery selector timeout.

Separate what to test from how to execute it. The test-plan builder lets a non-engineer describe a flow in plain English. A translation layer turns that description into a sequence of intents (navigate, select, verify) that the execution engine carries out against the live site, with the agent making moment-to-moment decisions about how to handle whatever is actually on the page. A separate reporting service aggregates results, evidence and diffs across runs so failures are traceable to a specific step, not just a red X.

Keeping test intent separate from page structure is what made the suite survive redesigns. A layout change didn't break the test; the agent adapted its execution the same way a human tester would have, and the plain-English test itself never needed to change.

Built to be trusted with production traffic. Because tests run against the live site, not a sandbox, false positives were the biggest adoption risk. Every run produces full evidence, so a flagged failure could be verified in seconds instead of triggering a re-run-and-hope cycle. Configurable environments (dev, staging, prod) meant the team could run daily production validation without risking real transactions, and rate limiting kept test traffic from skewing the site's own analytics.

The team went from one engineer owning QA to the whole team writing and reviewing tests, since no one needed to know code to add a flow.

The team went from one engineer owning QA to everyone writing and reviewing tests, and regression testing stopped being the release bottleneck within the first month.