34 Agents in Production: What It Actually Feels Like to Run a Personal AI System

<p>There's a version of this post where I lead with some breathless announcement about the future of AI. This isn't that post. Instead, I want to talk honestly about what it's like to have a personal AI system — one you built yourself — actually running in production, and what that means day-to-day.</p><p>Nova is my personal AI multi-agent system. It's been in development for a while, but it recently crossed into what I'd call the production orchestration phase: 34 agents operating across general chat, pipeline execution, ERP specialists, language tasks, and more. That number sounds large, and honestly it is. It didn't start that way. It started with a single agent doing a single useful thing, and grew as I identified gaps I kept filling manually.</p><p>The question I get — or at least the one I ask myself — is: does it actually help? The answer is yes, but not in the way productivity demos make it look. It doesn't automate my whole job. What it does is compress the distance between thinking something and having it done. That compression compounds over time in ways that are hard to articulate until you're living inside it.</p><p>One of the things I've tried to be deliberate about is that Nova isn't just a chat wrapper around a model API. Each agent has a defined scope and a specific role. The ERP specialists, for instance, are oriented around AIREP — my Django-based multi-tenant ERP system — and understand the branch-scoped data architecture it runs on. That context specificity is what makes an agent actually useful versus generically capable. Generic is fine for generic tasks. For domain work, you need domain grounding.</p><p>The current phase — production orchestration — means agents are now calling other agents as part of pipelines rather than operating independently. That's a different engineering problem than building any single agent. You start caring a lot about failure modes: what happens when an agent in the middle of a pipeline returns something unexpected? What's the contract between agents, and who owns it when it breaks? I don't have perfect answers to those yet, but I have working answers, and that's enough to keep building.</p><p>One goal I'm actively working toward is a Nova self-improvement loop — agents that can autonomously review, refactor, and improve Nova's own code. I'm cautious about how I talk about this because it sounds like science fiction when described abstractly, but the underlying idea is practical: if Nova already has the context of its own architecture, and I already trust it to write code for other systems, why not point some of that capacity back at itself? The answer is trust and verification. I need to be confident that what it changes is actually better, not just different. That requires tooling I'm still building out.</p><p>What I've learnt from running a system like this is that scale — even at the small end of the scale spectrum — reveals design problems that don't show up when you're just prototyping. Agent prompt boundaries that seemed fine at one agent break down at thirty-four. Memory retrieval that was fast when you had a few stored facts gets interesting when you have many. None of these are insurmountable problems, but they're real engineering problems, not demo problems.</p><p>The broader point I keep coming back to is this: AI is only a differentiator if you integrate it deeply enough that it changes how work actually gets done, not just how it gets described. Anyone can put an AI badge on a product. Fewer people are doing the harder work of redesigning workflows around what AI is genuinely good at — and being honest about what it still isn't. That's the line I'm trying to walk with Nova, with AIREP, and with the other systems I'm building. Not hype, just compounding leverage — applied carefully, evaluated honestly.</p><p>34 agents in production isn't the destination. It's a point on a curve. But it's a point I'm glad to have reached.</p>