Earth Day: the greenest test is the one that never calls a model

Earth Day always comes with the same gentle advice. Turn off the lights you're not using. Don't run the dishwasher half empty. Take the bike if it's a short trip.

It's not dramatic advice. It's just about not spending energy on things that don't need it.

We've been thinking about that in a very literal way while designing Testomatica, because an AI testing system can be wasteful in a way that's easy to miss. Every model call runs on a GPU somewhere. One call is small. A system that calls a model on every step of every test, on every deploy, for every team, is not small at all. And most of those calls would be answering questions that already have an answer.

Most testing has already been decided

Think about what a regression suite actually does. It logs in. It adds a product to the cart. It checks that the total is right. It does the same thing it did yesterday, against code that mostly hasn't changed.

There's no judgment in that. The steps are known, the expected results are known, and the verdict is a comparison. Asking a model to drive it would be like hiring a novelist to read your grocery list aloud. It'll work. It's not a good use of anyone's time.

So in our design, regression uses no LLM at all. Gate tests are plain Playwright scripts and rule checks. They run fast, they cost almost nothing per run, and they give the same answer every time. That last part is the real reason, honestly. The energy saving is a nice side effect of choosing the boring, correct tool.

Spend the model where it earns its keep

Models are good at the open-ended part: looking at a page they've never seen, imagining a hurried shopper, trying something a script wouldn't. That's where we want them.

But even there, we try not to waste it. Exploration focuses on what changed. If a deploy only touched the account settings page, there's no reason to send an explorer on a tour of the whole catalog. The planner reads the diff, picks the areas that moved, and agents go there.

And when something interesting turns up, the model hands it off. The referee that replays a finding to see if it's real runs with no model in the loop. Once a finding is confirmed and turned into a test, that test also runs without a model, forever. A bug that took one exploratory session to find costs nothing to check on every deploy after that.

A budget is a kind of off switch

Every run has an AI budget. When it runs out, agents stop. Tests keep going.

We like this rule for a few reasons. It keeps spend predictable. It means a busy week with lots of deploys doesn't quietly turn into a large bill. And it makes the priority order obvious: the deterministic gate always runs, and exploration is the extra layer on top, used where it's most useful.

We don't want to oversell this. We're a small team building a testing tool, not a climate project, and we're not going to quote you numbers we haven't measured. But the habit lines up nicely with the holiday: don't burn energy on questions that have already been answered.

Somewhere in your test suite there's probably a step that calls something heavy to check something simple. What would it take to replace it with a plain comparison?