This is Part 1 of AI in QE — a series on how we're using AI to remove toil across the quality engineering stack.

QE teams aren't slow because engineers lack judgment. They're slow because engineers spend most of their time on work that doesn't require any.

API test coverage has always carried a backlog. Not because the work is hard. Because it's tedious. Writing test cases for a microservice with no contract docs, hunting test data across domains, triaging CI failures that may or may not be real. This is where SDET hours disappear. Quarters pass. Backlogs compound.

We built a shared testing platform: 15 to 20 BDD steps covering GraphQL, REST, and gRPC. Intent-driven by design. Steps declare what's being tested, not how the protocol works. Heavy computation in the background. Minimal surface area for engineers.

Then we wired it to an AI agent via MCP.

An engineer gives a minimal prompt: service, operation, scenario. The agent generates a Gherkin feature file using predefined SDK steps. A companion DataGen service generates test data automatically. No manual domain lookups. No waiting on contract docs.

flowchart TD
    A[Engineer] -->|minimal prompt| B[AI Agent via MCP]
    B -->|selects steps from| C[Shared Step Library\n15-20 BDD steps\nGraphQL · REST · gRPC]
    B -->|generates| D[Gherkin Feature File]
    E[Test Data Service] -->|auto-generates test data| D
    D -->|executes against| F[Target Microservice]
    F --> G[Test Results and Report]

API test generation that took days, often queued for quarters, now takes minutes. Engineers don't need to know the SDK internals to produce coverage. The agent knows the step library. DataGen knows the data relationships. The engineer decides what to test.

The design that made it possible: 15-20 simple, declarative steps with heavy computation behind them. When the surface area for AI is clean and structured, generation quality is high. The SDK design is what makes the prompt work. Not the model.

If you're building AI-enabled quality tooling, start with the step library. The AI is only as good as what it has to work with.