For most of computing history, bespoke software was a luxury. If you wanted a tool built exactly for how you worked — your workflows, your data sources, your quirks — you needed either engineering time or enough money to buy it. Everyone else used off-the-shelf software and adapted themselves to it, which is a strange inversion: humans reshaping their habits around the constraints of software rather than the other way around.
AI coding agents changed this equation in a way I don’t think we’ve fully absorbed yet.
I’ve been building a small ecosystem of personal tools over the past few months — a background job runner for AI agents, a CLI for querying my transit system, a tool that checks my health data and writes a weekly summary, another that monitors regulatory circulars relevant to my work. None of these are products. None of them needed to be. They’re just small, precise solutions to friction I kept encountering, built in the gaps between other work, maintained by the same agents that built them.
The shift isn’t that these tools are easier to write (though they are). It’s that the cost of specificity has collapsed. Off-the-shelf software is generic because genericity scales — one codebase, millions of users. But when the cost of building something custom approaches zero, the calculus flips. Why adapt to a generic tool when you can have one that fits exactly?
The interesting question is what this means at scale. When a single person can maintain a fleet of small, personalised tools — each purpose-built for a specific need, each co-evolved with its owner through ongoing conversation with an agent — you get something that looks less like software and more like an extension of how that person thinks. The tool knows which repositories matter to you, which topics you’re tracking, how you like to receive information, what cadence makes sense for your life.
This is meaningfully different from personalisation in the current sense — the kind where an app learns your preferences through a recommendation engine or surfaces content based on your history. That’s personalisation as inference about you. What I’m describing is personalisation as authorship by you. The tool doesn’t guess at your needs; you specify them, refine them, and evolve them through the same conversational interface you use to ask questions.
There’s a risk worth naming. Systems built for one person are brittle in ways that generic systems aren’t — they carry assumptions that stop being true, accumulate cruft as needs change, and have no community of users to surface bugs. The maintenance burden is real, even when agents handle most of it. And there’s something to be said for constraints: generic tools force a kind of discipline, because you can’t keep adding personal exceptions to something other people also use.
But I think we’re in early days of understanding what personalised systems actually feel like to live with over time. The tooling is new enough that most of us are still in the building phase, not the maintaining phase. The harder questions — about coherence, about when to generalise versus specialise, about what happens when your needs change and your tools don’t — are only now becoming visible.
What I can say from where I’m standing: the experience of using a tool that was built for exactly how you work, by an agent that knows your context, is qualitatively different from using software you’ve adapted to. It’s not just faster. It feels like the friction was removed from a direction you didn’t know friction could be removed from.
That’s new. I don’t think we have good language for it yet.