34 Agents in Production: What Running Nova Actually Looks Like

<p>Nova is in production. Not "works on my machine" production, not "I demoed it once" production — actual daily-use, orchestrating-real-work production. It's running 34 agents across general chat, pipeline execution, ERP specialists, language tasks, and a few others I've bolted on as specific needs came up. I want to write down what that actually looks like, because most of the discourse around AI agent systems is either breathless hype or academic abstraction.</p><p>The honest version: it started as a personal productivity experiment and turned into something I now consider a core part of how I work. The 34-agent count sounds impressive until you understand that most of those agents are narrow and purpose-built. There's no single god-agent trying to do everything. There's an ERP specialist that understands AIREP's data model. There are pipeline agents that handle discrete, repeatable tasks. There are language agents for drafting and editing. Each one has a tight scope. That's intentional.</p><p>The orchestration layer is where the interesting problems live. Getting one agent to do something useful is relatively straightforward. Getting a system to route the right task to the right agent, handle failures gracefully, and return something coherent — that's the actual engineering work. I'm still iterating on this. The production orchestration phase doesn't mean it's finished; it means it's stable enough that I trust it with real tasks and it's integrated into real workflows. There's a difference between "works" and "done," and I've made peace with the fact that systems like this are never done.</p><p>One of the things I keep coming back to is the framing of AI as a tool versus AI as leverage. A tool is something you pick up when you need it and put down when you're finished. Leverage compounds. Nova isn't something I open when I have a specific question — it's woven into how I manage projects, how I think through architecture decisions, how I handle the administrative overhead of running multiple projects simultaneously. The compounding effect is real, but it takes time to materialise. You have to actually build the integrations, not just have access to an API.</p><p>The self-improvement loop is the next frontier I'm working toward. The goal is agents that can autonomously review, refactor, and improve Nova's own code. That's not a gimmick — it's a practical necessity. I'm one person managing AIREP, Find a Sign, Sweeper Parts, client work, and this system simultaneously. If Nova can shoulder some of its own maintenance burden, that's genuine leverage. If it can't, it's just another thing I have to maintain. The architecture decisions I'm making now are partly about keeping that future state achievable.</p><p>There's also something worth saying about why I built this myself rather than just using an off-the-shelf product. Partly it's because no off-the-shelf product fits the specific combination of things I need — deep integration with my ERP work, my project context, my clients. But partly it's because I think technical ownership matters. If I'm positioning AI as a core business competency, I need to actually understand how these systems work under the hood. Using a black box isn't a competency; it's a dependency.</p><p>The part that surprised me most in getting to this point is how much of the work is just software engineering. Prompt design matters, model selection matters, but the thing that actually makes or breaks an agent system is the same stuff that makes or breaks any software: clear interfaces, good error handling, sensible data flow, and ruthless simplicity in scope. The agents that work well are the ones with the tightest, clearest responsibilities. The ones that cause problems are the ones I tried to make too clever.</p><p>I'll keep writing about this as it evolves. There's a lot more to say about specific architectural decisions, about what's worked and what hasn't, and about where the production orchestration phase goes next. For now, the short version is: 34 agents, real work, still iterating, no regrets.</p>