Test driven development with AI Agents

What is TDD (Test Driven Development)?

At its core, TDD revolves around a cyclical process known as Red-Green-Refactor. The cycle commences with the “Red” phase, where a developer writes an automated test case that defines a desired feature or behavior. This test is intentionally written to fail initially because the corresponding code does not yet exist. This initial failure is critical as it ensures that the test is indeed testing the intended functionality and can catch errors once the code is implemented.

The subsequent “Green” phase involves writing the minimum amount of code necessary to make the previously failing test pass. This principle encourages developers to focus on the immediate requirement defined by the test, maintain a clean codebase, and prevent over-engineering.

Finally, the “Refactor” phase focuses on improving the structure, readability, and maintainability of both the test code and the production code, all while ensuring that all existing tests continue to pass. Refactoring ensures that the codebase remains healthy and adaptable to future changes without introducing regressions, with the existing test suite acting as a safety net during this phase.

How to Fully Automate your TDD pipeline?

There are a lot of factors affecting the performance of AI Agents while coding. From the LLM Model to how you structure your code and your development pipeline. We have found out TDD to be effective with Windsurf with Claude 3.5 sonnet. The following is a sample task that is implemented in TDD.

What do you need?

Before we start coding, we need the following:

Enough Tests

Make sure you already have tests according to TDD, and they cover most of the scope of your problem that you think is logical and helpful for the AI Agent. You don’t need to change or customize anything for sake AI Agent. If your tests and their naming convention are based on a standard, it would be better. As the first step, the AI Agent will see these tests to start implementing.

Here is an example of a test I have that tests whether a document has been correctly inserted into MongoDB:

Interfaces

An Interface in Java (or any other programming language) is a structure defining what the classes should look like and what methods they should implement. To guide the AI Agent even further, it would be helpful to create an Interface for your repository to insert documents:

Specific description of the task

Lastly, we need a specific task description. Usually, you can use JIRA or GitHub issues where you define the task. This is ours:

Start Vibe Coding

Vibe coding means you describe what you need in plain English (or even voice commands), and AI generates the code for you in real time. We go into detail on Vibe Coding in this blog. Here is the prompt I used in Windsurf to complete the task for me:

Implement the following query as a combination of one or more named queries and Java code.

Find all drivers who have completed at least X trips with a rating above 5 in a given date range and have never received a rating below 3 stars.

Note:
You do not need to find a single query to solve this task (you can use a combination of Java code and named queries), but you have to keep ORM-performance in mind, i.e., make sure that your solution is also reasonably fast if you have many entities. During the discussion session you should be able to explain what types of problems can arise with badly written queries.

@Ass1_2_2Test.java#L35-60 these are the corresponding tests. Its implemented in @DriverDAO.java#L34-63

Mentioning the files is the most important aspect when vibe coding. Now sit back and see the AI Agent do its magic. It implemented the Class, ran the tests, and reiterated to make the tests pass: