New year posts are usually predictions, and predictions are usually wishful thinking wearing a suit. So let me do it differently: I'm going to tell you what I'm building toward in 2026, and why. You can hold me to it in December. Some context for why I trust this particular hunch. By day I build agent systems at HD2.ai. And I've been in and around the music industry since I launched Flex Records in 1994. I have watched every technology wave hit that business: CDs, downloads, streaming, Web3, and now AI. And the pattern is always the same: the thing that wins is never the revolution. It's the boring version that actually fits into how people already work. With that lens, here's 2026.

One: agents stop being demos and become plumbing

2025 was the year agents became possible. 2026 is the year they become boring, which is the good outcome. The interesting question stopped being "can an agent do this task?" (mostly yes) and became "can it do the task inside my constraints, on my data, without me babysitting it?" That's a plumbing problem: protocols (MCP and friends), scoping, audit trails, graceful failure. The teams who win this year will spend less time demoing agents and more time giving them boring, well-scoped jobs that never make the news. That's when you know a technology has landed.

Two: the music business adopts agents the way it adopted streaming (late, then completely)

The industry is about to discover what indie builders already know: label operations are agent-shaped. Catalogue metadata is a forever-mess. Back-catalogue discovery, "which of these 400 tracks from 2003 are suddenly sync-worthy", is retrieval, not artistry. Royalty reconciliation is exactly the kind of tedious, rule-based work agents eat for breakfast. None of this touches the creative act, and that's precisely why it'll be adopted without the usual studio-versus-software war. The business around the music gets automated; the music doesn't.

Three: local context beats bigger context windows

The past two years were about cramming more tokens into prompts. The next phase is retrieval and local context: agents that read your files, your catalogue, your logs, and stay on your hardware. For anyone with proprietary data (and a label's catalogue is nothing but proprietary data), an agent that never phones home beats a smarter model that does. I'll be writing more about this stack as the year goes on.

Four: sovereignty stops being paranoid and becomes procurement

On-prem and self-hosted AI used to be a stance. In 2026 it's a line item. The questions "where does this data live" and "who else can see it" are becoming standard procurement checkboxes, and the builders who can answer them cleanly will win work the cloud-only crowd won't even see.

A quieter take

Here's the resolution I'd offer instead of predictions: in 2026, own your intelligence layer. Whatever your version of the business is (codebase, catalogue, client list), the compounding advantage isn't the model you rent; it's the context, tooling, and workflow you own around it. That's what I'm building toward this year. Ask me in December how it went.