Strict TDD with AI agents: slower, pricier, and better

· Published · 1 min read

Have you found a good setup to make your agents do TDD?

I am running lots of experiments at the moment.

One of them: forcing agents to work in strict TDD (the real thing, mind you, not test-before).

The result

So far, I find the end result is almost always better:

  • a better design
  • problems spotted early
  • much sturdier regression tests

The trade-offs (because there are some): slower agents, and higher costs.

Which suits me just fine, because I work with ~5-6 agents in parallel… and I would rather they take their time and do better than jump from one blunder to the next.

Going further than the prompt

In short, I want to dig into the TDD + agents direction.

I have already used the TDD skills from Superpowers and from Matt Pocock.

They are great, but they are still… just a prompt.

So I was wondering whether better setups exist to go further. Through hooks, for example?

Have any of you tried? Any lessons learned? If this rings a bell for anyone, I would love to hear about it!

Sources