I’m spending the summer taking some time off work, and I can’t tell you how pleasant it is. The pace of life has slowed, and I highly recommend it if you ever get the chance.

In between ticking off a long list of house jobs, I’ve had the rare opportunity to get my hands back on the keyboard and play with code. It’s been years since I last had the space to do this, and I’d forgotten how much I enjoy it. Who knew that improving test coverage on a legacy codebase could feel anything but soul-destroying? Especially when you’ve got AI coding assistants doing the heavy lifting.

But this isn’t really about my summer of leisure. It’s about what that tinkering time has surfaced for me: some thoughts on Agentic AI and, to a lesser extent, Robotic Process Automation.

If you haven’t yet been drawn into conversations about how Agentic systems could transform efficiency, you will be soon enough. On paper, the potential is compelling: autonomous agents stitching together APIs, orchestrating workflows, and reducing repetitive human effort. But I can’t shake the concern of what happens next.

Take the codebase I’ve been working on. It’s solid, it does what it’s meant to, and it’s been running in production for years without much fuss. But even at just eight years old, it’s already becoming harder to support without significant investment in modernising its design patterns. Without that ongoing care, critical dependencies like database versions will move on, leaving the platform stranded.

At LDC Via, we’re big enough nerds that this kind of upkeep feels normal, and we’re not especially constrained by budgets. But in a large enterprise, I can easily imagine the temptation to adopt the new and shiny at the expense of maintaining the old and forgotten.

And here’s the risk: as you experiment with Agentic AI, even if you’re hooking your agents up to a rock-solid, API-driven backend, what happens next year? Or in five years?

Who’s going to know how that API works?

Who’s going to make sure it still works?

And if no one does, what happens when your “autonomous” agents quietly fail?

This isn’t an argument against Agentic AI, far from it. The potential is huge. But like so much in tech, its success will depend not just on what we build, but on what we’re willing to maintain.