Nova Is in Production. Here's What That Actually Means.

<p>Nova is my personal AI system. I've been building it alongside everything else — AIREP, Find a Sign, client work — and for a while it lived in that comfortable zone of being impressive enough to demo but not something I actually depended on. That changed. It's now in what I'm calling its production orchestration phase, and I want to write honestly about what that means and what it doesn't.</p><p>The system currently runs 34 agents. They span general conversation, pipeline execution, ERP specialists, and language tasks. That number sounds like a flex, and I'll admit it reads that way on paper. But the more useful framing is this: the system has grown large enough that I can no longer hold all of it in my head at once. Which means it's either production, or it's just complicated. The goal is to make it the former.</p><p>"Production orchestration phase" is my term for the point where the system is doing real coordination work — agents handing tasks to other agents, pipelines running without me shepherding every step, outputs I actually use rather than just evaluate. It's not a marketing milestone. It's an internal threshold I set for myself because I needed a way to distinguish "this works in a test" from "this is doing something." We're past that line now.</p><p>What I've learned building to this point is that agent systems don't fail the way you expect. Individual agents are usually fine. The failure modes live in the handoffs — mismatched context, an agent that assumes it has more state than it does, a pipeline that works end-to-end in isolation but quietly breaks when another part of the system changes. This is the same class of problem as microservices, and the solution is the same one you learn there too: explicit contracts, narrow interfaces, and a lot of logging.</p><p>The other thing I've learned is that "personal AI" is a deceptively hard problem because the feedback loop is tight in the wrong direction. When I build something for a client, there's distance — I can evaluate their experience somewhat objectively. With Nova, I'm the user, the developer, and the QA process. It's very easy to build something that works for how I think I work rather than how I actually work. The moments where Nova surprised me by doing something genuinely useful that I hadn't anticipated — those are the signal. Everything else is me building a very elaborate mirror.</p><p>One of my active goals is to build a Nova self-improvement loop: agents that can review, refactor, and improve Nova's own code autonomously. I haven't shipped that yet. What I have done is get to a state where the infrastructure to support it is real rather than aspirational. The orchestration layer exists. The agents can be tasked. The plumbing is in place. The self-improvement piece is the next hard problem, and it's a genuinely different kind of hard — less about engineering and more about how you define "better" for a system you interact with every day.</p><p>The broader point I keep coming back to is this: AI isn't a feature I'm adding to my projects. It's the primary leverage point across everything I'm building — AIREP, Find a Sign, Nova itself, the client work. The projects that are going to matter in three years are the ones where AI is load-bearing, not decorative. Nova is my proving ground for that thesis. It runs on real tasks, against real deadlines, and when it fails I feel it directly. That's the right kind of production.</p><p>I'll keep writing about this as it develops. The honest version of building something like this is more interesting than the polished retrospective, so I'd rather document it as it happens — including the parts that don't work yet.</p>