Agentic Test Engineering for a Payments Platform

A global payments provider ran end-to-end regression for its business and admin banking portals largely by hand. The expensive flows — ACH batch initiation with two-level approval, extended-cutoff batch processing, check-image retrieval — each span two portals, a scheduler service, and entitlement setup before the first assertion runs. Writing one reliable test took days, and every UI change put the suite back into maintenance.

• Structured Markdown test plans with explicit SETUP preconditions, numbered steps, and expected outcomes — readable by business analysts, executable by agents, versioned with the code.
• Three declaratively defined agents: a test planner (requirement to plan), a test generator (plan to Playwright spec), and a test healer that debugs failures using live browser snapshots, console output, network traces, and generated locators.
• Model Context Protocol integration so agents drive a real browser session rather than guessing selectors from static code.
• A shared framework layer — seeded data, a reusable common package, an execution wrapper, an automated summary reporter — wired into CI via GitHub Actions.

Regression scenarios that existed only as tribal knowledge became executable, versioned specifications. New scenarios start from a plan rather than a blank file, and failing tests arrive with browser evidence attached instead of a stack trace.

Now the reference automation pattern across the client's architecture practice