I think we'll look back on 2025 as a dividing line. Projects started before it carry the fingerprints of human hands all over them. Projects started after are largely—sometimes completely—built and architected by LLMs.
That distinction matters more than you'd think.
A question came up in the Builder Methods Pro community this week: what defines a "legacy" codebase?
To me, it's simply anything beyond a brand new MVP. The core functionality has been established. Now the work is a mix of maintenance, bug fixes, and building on top of what's already there.
But 2025 added a new dimension to what makes code "legacy." I'm seeing it clearly in my own projects.
Codebases started in the pre-AI or early-AI days are harder to maintain with agents. Changes are slower and more comborsome. But codebases built recently with agents from the start? Those are much easier for agents to pick them up and work on.
This isn't an an LLM thing. It's a human thing.
It's about where we put our personal imprint on our craft. In the pre-AI days, we poured our thoughts, opinions, and design patterns directly into the code itself. But in the AI era, our imprint shifts to the planning phase. We leave the actual code writing to the agent.
That's not carelessness—it's the opposite. We're now able to put our creative energy into where it matters most: the spec. And because we were never precious about the code in the first place, agents can work on it without friction. No hidden context to decode. No sacred architecture to tiptoe around.
But those pre-AI codebases are still (very much) in play. So how do we get them into a workable state for the AI era?
That question inspired my latest update to Agent OS v3, which I released this week.
Standards, as a concept in Agent OS, help you capture your knowledge, design patterns, and decisions that aren't obvious from an agent's scan of your codebase. V3 now includes a tool for automatically discovering and documenting your standards for you. Once in place, it makes you (and your agents) much more confident maintaining those aging codebases in 2026.
