Ideas
Brownfield vs. greenfield is the wrong debate in the age of AI
Brownfield or greenfield predicts little: what matters is whether an agent finds the right context within clear boundaries, from day 1 to day 1,000
People talking about brownfield vs. greenfield in the age of AI are missing the point.
AI doesn’t care when a project was created. Whether it started yesterday or 20 years ago is not what matters. What matters is the size of the context AI needs to work effectively.
The usual objection I hear is: “But a greenfield project is necessarily small. It starts with zero lines of code.”
Fair enough. But two months and a pile of features later, that same project can look just like any other codebase. Has it become brownfield overnight?
The real issue isn’t the age of the project. It’s how much context AI needs to understand it and make the right changes.
A well-structured project, with clearly defined responsibilities and properly bounded contexts, can fully benefit from AI, whatever its age. A project that ignores software engineering best practices and domain boundaries will be much harder for AI to work on.
AI shouldn’t be a disposable tool that works wonders in a project’s first three days and then becomes less useful as the codebase grows. It should be part of the entire development lifecycle and help prevent architectural drift as the project evolves.
That’s where the context budget comes into play. More on that here: Plus le projet grossit, plus l’IA se trompe.
There’s also a question of hygiene in how you write skills, structure
CLAUDE.md/AGENTS.md files, and give AI its context.
Redundancy is not a feature. If you’re working on a PHP codebase, there’s no need to keep reminding the agent that it’s PHP. It already knows.
In the age of AI, brownfield vs. greenfield is a poor predictor of how effective AI-assisted development will be. What matters is how well a codebase lets an agent discover the right context, reason within clear boundaries, and make changes safely. The real challenge is preserving those properties as the project grows.
Don’t optimize a codebase for AI only on day one. Engineer it so that AI can keep working effectively on day 1,000