Verifying AI-Generated Software with Approved Scenarios

How can you trust the code an AI agent produces? One of the important ways is through automated acceptance tests. In this session you will learn about Approved Scenarios, a form of executable feedback that you use as part of the development process to steer an agent’s work. It’s an approach that has been around for many years and is particularly useful in legacy code. It now makes even more sense, when generating code is cheap, but the consequences of mistakes and bugs are just as serious as ever.

What you work on

In this masterclass you work on a small example where the code has some functionality but inadequate tests. You steer a coding agent to create the needed Printer and Parsers and some initial test cases, then develop further, getting visual feedback on progress and key functionality. The emphasis is on working effectively with AI agents on existing code where the domain is complex and bugs have significant consequences.

Who it is for

Approved Scenarios are normally used on larger pieces of code, often as an alternative to something like Cucumber, since you can use them in a BDD or TDD-like process. The masterclass is not directed at testing professionals. The focus is on helping developers to use Approved Scenarios and related techniques to get control over legacy code when using coding agents.

See also approvaltests.com and texttest.org.

After the masterclass you can

• Steer agentic AI using Approved Scenarios in order to keep control over functionality and quality in an existing codebase.
• Describe the characteristics of Approved Scenarios compared with ordinary unit- or assertion-based testing in a TDD-style process.
• Compare Approved Scenarios to other kinds of acceptance tests, for example those using Cucumber.
• Use Approved Scenarios as characterization tests in legacy code as part of design improvement using agentic AI.

Practicalities

The masterclass runs from 09:00 to 16:00 on 25 February. Bring a laptop and expect to work on coding exercises that illustrate the purpose and practice of Approval Testing. An online programming environment is provided; alternatively you can download the exercises from GitHub and work on them locally. You also need to access Miro online activity boards from your laptop.

Code examples are in Java, but the principles and approach apply equally to C#, C++ and other mainstream programming languages.