What Running a 34-Agent AI System in Production Actually Looks Like

<p>There's a version of this post where I tell you Nova is live, everything works, and AI has changed my life. That's not this post.</p><p>Nova is my personal AI system — a multi-agent architecture I've been building alongside everything else: AIREP, Find a Sign, Sweeper Parts, client work. As of now it's in what I'm calling the production orchestration phase, which means the system is real, it's running, and I'm using it daily. It currently spans 34 agents across general chat, pipeline execution, ERP specialists, language tasks, and more. That number sounds impressive until you start operating it and realise that orchestrating 34 things is its own full-time problem.</p><p>What I mean by production orchestration is simple: the agents exist, they have defined roles, and the plumbing to route tasks between them is in place. What I'm working through now is the harder layer — making sure the right agent gets the right task, that context is passed cleanly, that failures don't silently disappear, and that the whole thing compounds into something more useful than a collection of isolated tools. That last part is the goal. Isolated tools are fine. A system that compounds is different.</p><p>The reason I started building Nova instead of just using one of the commercial AI products is that I wanted AI to be a genuine structural advantage — not a subscription I tab into when I'm stuck. If AI is your primary leverage point in software engineering and business operations, you can't afford to treat it like a search engine with better grammar. You need it integrated into how you actually work: understanding your projects, your decisions, your context. No commercial product does that out of the box, and most of them aren't designed to.</p><p>Nova is the backend AI service visible on keirantrace.com, but more practically it's the system I use to manage context across everything I'm running simultaneously. AIREP has its own ERP specialists inside Nova. There are agents that understand the Find a Sign marketplace model. There's context about Sweeper Parts, about client work, about open architectural decisions I'm sitting on. The goal is that when I sit down to work, I'm not re-establishing context from scratch every time — Nova holds it.</p><p>One of the goals I've written down is building a Nova self-improvement loop: agents that autonomously review, refactor, and improve Nova's own code. That's not running yet. I want to be clear about that. It's a goal, not a current capability. Getting the orchestration layer stable is the prerequisite, and that's where I am now. You don't build a self-improvement loop on top of a shaky foundation — you end up with something that confidently makes things worse.</p><p>The other thing I've learned from this build is that agent count is a vanity metric. Thirty-four agents sounds like a lot. What matters is whether the system reduces cognitive load or adds to it. Right now it's somewhere in between, which is honest. The parts that work well — memory retrieval, task routing for well-defined pipelines, context injection — save real time. The parts that are still rough require babysitting, which defeats the purpose.</p><p>I think the right mental model for a personal AI system at this stage is: you're building infrastructure, not using software. The payoff is compounding and delayed. The cost is real and upfront. That's fine if you understand it going in. It's frustrating if you expect it to work like a product on day one.</p><p>The production orchestration phase for Nova is about closing the gap between infrastructure and leverage. I'll keep writing about what that looks like as it develops — the wins, the failures, and the architectural decisions that turn out to matter more than they seemed at the time.</p>