Nova Is in Production: What 34 Agents Actually Means in Practice

<p>A few months ago Nova was a project. Now it's infrastructure. That shift — from "thing I'm building" to "thing I depend on" — happened gradually and then all at once, which is how most meaningful transitions in software seem to go.</p><p>Nova is my personal AI multi-agent system. It currently runs 34 agents across a handful of domains: general chat, pipeline execution, ERP specialists for AIREP, language tasks, and a growing set of agents that exist purely to coordinate other agents. That last category is the one I find most interesting, and most honest about what "production orchestration" actually means.</p><p>When people hear "34 agents" the instinct is to imagine some elaborate autonomous machine running loose. The reality is more grounded than that. Most of those agents are narrow and deliberate. An agent that understands AIREP's branch-scoped data model isn't general intelligence — it's a focused context carrier that lets me ask questions about multi-tenant ERP architecture without re-explaining the entire system every time. The value is accumulated context and specialisation, not magic.</p><p>What changed when Nova moved into production is that I stopped treating it as optional. It's now in the loop for real work: drafting, researching, reviewing code, managing reminders, and helping me context-switch across AIREP, Find a Sign, Sweeper Parts, and client work without dropping threads. When you're running four active projects simultaneously, the cognitive overhead of context switching is a real cost. Nova absorbs a meaningful chunk of that cost.</p><p>The architecture I've landed on is a pipeline model. Tasks come in, get routed to the right agent or chain of agents, and outputs feed back into the system or surface to me directly. Some pipelines are simple — one agent, one task. Others are multi-step: a brief goes through a research agent, then a drafting agent, then a review pass. I don't expose all of that to myself in real time. I just get the result. That abstraction is intentional. If I'm watching every intermediate step, I haven't actually offloaded anything.</p><p>One of the current goals I'm working toward is a Nova self-improvement loop — agents that can autonomously review and refactor Nova's own code. That's not live yet, and I won't claim it is. But it's the logical next step from where the system sits now. Once you have a reliable orchestration layer, pointing some of that capacity at the system itself is the obvious move. The compounding effects of a system that can improve its own tooling are significant, even if the early iterations are mostly catching obvious inefficiencies rather than doing anything clever.</p><p>The reason I'm investing this heavily in a personal AI system rather than just using off-the-shelf tools is competitive. AI isn't a feature I'm adding to products — it's the primary leverage point in how I build and operate. AIREP, Find a Sign, Nova itself: the goal is for AI to be a core differentiator in each of them, not a wrapper around someone else's product. That requires actually understanding the systems, not just prompting them.</p><p>Building Nova has taught me more about practical AI integration than any tutorial or course could. The failures are instructive in a way that reading about failures isn't. You learn where context windows actually become a problem. You learn which tasks agents handle well and which ones they confidently handle badly. You learn that the hardest part of a multi-agent system isn't the agents themselves — it's the routing logic and the memory architecture that lets agents accumulate useful context over time rather than starting from zero each session.</p><p>Production is a good word for where Nova is now. Not because it's perfect, but because I'm relying on it. That's the real threshold.</p>