Yes. To some degree, but AI can create the unit tests and benchmarks as you create the tasks.
The workflow I’m using is I create a ticket for feature X, then the AI quizzes me about the feature until we have all the details established - this is a great brainstorming tool by itself as it often makes me aware of things I hadn’t considered. All the details are added to the ticket.
The AI also extracts important parts from this discussion that are relevant to the overall architecture of the program and stores these in a specification document. This builds up over time and can be revised as the program changes and grows.
If necessary the ticket is then broken down into sub-tickets with different tasks, like researching, prototyping, implementing, etc.
After that it implements the tickets, which includes creating unit tests and benchmarks if necessary (the details will have been agreed upon in the quizzing stage).
After the implementation it runs a review pass, comparing the code to the spec and the tickets to make sure everything that was agreed was actually done. If there are mistakes it will ask you what you want to do about it and usually the solution is to run a fix pass.
How the tests and benchmarks are used varies depending on what you want. I have a script that runs clang-tidy and clang-format, all the tests, and then the AI checks that the benchmarks are being hit. I’ll also do profiling now and again and larger codebase reviews.
This workflow comes mostly from Matt Pocock. I’m using his AI skills files with Claude, and I’m using Herdr to manage the agents, which is apt because I feel like I’m an AI shepherd now. Here’s a video explaining the process: https://youtu.be/M6mYodf0dJM?t=504
Additionally I created a delegate skill which lets my Claude session run multiple workers or agents in parallel in separate worktrees and then merges and tests them. This makes it much faster to implement a complex feature that has been split into sub-tickets.