MCP-Powered QA Agents: Evaluation Criteria and Five Companies to Explore

The article examines how AI agents using Model Context Protocol can run software tests, stressing that controlled actions and verifiable conclusions matter as much as the underlying model. In the described architecture, an MCP server exposes approved testing operations such as Playwright browser tests and API checks, and enforces permissions so authorization stays in application code rather than prompts. It recommends separating observed failures from hypotheses, allowing an "unknown" diagnosis when evidence is thin, logging tool activity, and requiring human approval for sensitive actions. The piece also lists five firms offering relevant QA services or platforms: GeekyAnts, Accenture, IBM, TCS, and Capgemini (via Sogeti), while noting that inclusion does not imply equivalent capabilities. It suggests pilots in bounded environments with known failures, measuring diagnosis accuracy, time to actionable findings, and cost.