What Running 34 AI Agents in Production Actually Looks Like

<p>Nova is my personal AI system. Not a side project, not a demo — it's the operational backbone I use every day across client work, my own products, and general problem-solving. As of right now it runs 34 agents covering general chat, pipeline execution, ERP specialists, language tasks, and more. I'm calling this the production orchestration phase, which sounds grand, but what it actually means is: the system is live, it's doing real work, and I'm responsible for it when it breaks.</p><p>The number 34 sounds like a lot. It is. But the count is almost beside the point — what matters is the architecture underneath it. Each agent has a defined scope. There's no single monolithic AI that does everything; instead, there are specialists that get invoked when the task matches their domain. An ERP-focused agent handles AIREP queries differently to how a general chat agent handles a question about a client's website. Routing is the hard problem. Getting the right task to the right agent, with the right context, without the user needing to think about it — that's where most of the design effort goes.</p><p>One of the things I keep coming back to is the difference between AI as a tool and AI as infrastructure. A tool is something you pick up when you need it. Infrastructure is something you build on top of, something that has to be reliable, observable, and maintainable. Nova is the latter. That changes how you think about it. You stop asking "what can this AI do for me today?" and start asking "what does this system need to keep running well?" That means thinking about failure modes, context management, agent boundaries, and — increasingly — how agents can improve the system itself.</p><p>That last part is something I'm actively working toward: a self-improvement loop where agents can review, refactor, and improve Nova's own code. It's not science fiction — it's a logical extension of what the system already does. If an agent can help me write better Django views, there's no fundamental reason it can't eventually flag its own poorly structured prompts or suggest tighter tool definitions. The challenge is trust and verification. You need enough observability to know when an autonomous change is an improvement versus a regression. I don't have that fully solved yet, but the goal is concrete.</p><p>Running this in production has also sharpened my thinking on AI integration in my other products. AIREP, my Django-based ERP system, is being built with AI as a first-class concern — not bolted on after the fact. The multi-tenant architecture means each client's data is branch-scoped and isolated, which creates real constraints around what context you can safely pass to an AI layer. You can't just dump a user's query into a model with no regard for data boundaries. Getting that right is painstaking, but it's the kind of work that separates a system people can trust from one that just demos well.</p><p>Find a Sign, my Australian signage marketplace, has a different AI angle. The core value proposition there is transparent, customer-first discovery — no pay-to-rank, no supplier manipulation. AI fits into that as a better matching layer, not a monetisation lever. I'm wary of the pattern where AI gets used to obscure commercial incentives behind a veneer of "smart recommendations". That's the opposite of what I'm building toward.</p><p>The honest summary of where I'm at: Nova works, it's useful, and it's taught me more about building reliable AI systems than any tutorial or paper could. The 34 agents aren't impressive because of the number — they're the result of iterating on real problems, cutting agents that didn't pull their weight, and refining the ones that did. Production has a way of making you honest about what's actually working.</p><p>The next phase is less about adding more agents and more about making the system smarter about itself — better routing, better context handling, and eventually that self-improvement loop. I'll write about how that progresses as it does.</p>