Web Directions Year round learning for product, design and engineering professionals

Building Frameworks Building Systems

Imagine needing to build hundreds of interactive science and maths games at scale. Not one game that's beautiful and perfect. Hundreds. Each one tailored to a specific learning outcome, a particular age group, a unique pedagogical approach. The traditional response — hire more engineers, write more code — hits a wall fast. You'd need a team that grows with the number of games you want to ship.

Curriculum Associates faced exactly this problem. The solution wasn't more engineers. It was infrastructure that builds the games for you.

This is where the idea of treating AI as a build step becomes concrete and powerful. Rather than AI as a coding assistant that engineers use interactively, AI becomes part of the delivery pipeline itself. You define the specs — structured descriptions of what the learning game needs to accomplish, what interactions it should support, what concepts it should cover. The pipeline takes those specs and generates tested, deployable experiences. The AI isn't helping an engineer code; it's being the coder in an automated workflow.

The implications run deep. First, it decouples domain expertise from coding expertise. You don't need a game programmer to specify a game anymore; you need someone who understands mathematics education or science learning, who can write a spec. That person might never write a line of code. The infrastructure translates their intent into a working system.

Second, it creates consistency at scale. When humans code each game individually, you get variation: this one handles errors elegantly, that one has a clunky UI, another has a great learning arc but poor performance. When a pipeline builds them, you set the quality baseline once and every game meets it. Your floor is higher.

Third, it enables experimentation. If generating one game takes hours, you can only try a few approaches. If it takes minutes, you can run dozens of variations, test them with actual learners, measure outcomes, and iterate. The cost structure of exploration changes fundamentally.

There's a catch, of course: the specs have to be precise enough that AI can translate them into working systems. That means investing heavily in how you describe what you want. Vague specs produce vague games. But that constraint is actually valuable — it forces clarity. When you have to be explicit about what makes a good learning experience, you learn things.

The technical challenges are real too. The CI pipeline needs to not just generate code but test it, validate that the generated game actually teaches what it's supposed to teach, catch regressions, and handle failures. You can't have a learning game ship with a bug any more than you could have a production system ship with one. The stakes are just different.

This approach has a name in the broader AI world: agentic workflows. Instead of AI being a tool humans use, it's a component in a system that accomplishes a goal. Instead of "AI helps me write code," it's "my system uses AI to write code as part of delivering value." The difference is subtle but shifts everything about how you architect, how you validate, and how you scale.

For organizations sitting with the question "how do we do more with the same team," this offers a radically different answer: don't ask your team to do more. Build systems that do it for you. Use AI not as a faster way for humans to code, but as the execution layer in a delivery pipeline designed around the outcomes you actually care about.

Ally Macdonald, a Staff Builder at Curriculum Associates, will explore this approach and the systems that made it work at AI Engineer Melbourne 2026 on June 3-4.