Nova Is in Production. Here's What 34 Agents Actually Looks Like.

<p>Nova is in production. Not "I have a chatbot running" production — I mean a multi-agent orchestration system with 34 agents handling different domains of my work and life, routing tasks between them, and operating without me holding its hand for every step. That milestone deserves a proper write-up, because the gap between "I'm building an AI system" and "my AI system is running" is wider than most people expect.</p><p>Let me be specific about what 34 agents actually means, because the number is easy to throw around and hard to make concrete. The agents are grouped by function: general conversation and reasoning, pipeline execution (long-running multi-step tasks), ERP specialists that understand AIREP's data model, and language tools for things like writing, summarising, and drafting. Each agent has a narrow responsibility. None of them do everything. That's the point.</p><p>The architecture decision to keep agents narrow was deliberate and took some discipline to stick to. Early on there's a temptation to build one very smart general agent and call it done. It's faster to prototype and easier to explain. But a general agent that does everything is also an agent that does nothing particularly well, and more importantly, it's an agent that's impossible to improve incrementally. When something breaks or produces bad output, you can't isolate the problem. When you want to improve one capability, you risk degrading another. Specialisation is the only path to a system you can actually maintain and trust.</p><p>"Production" for a personal AI system also means something different than it does for a SaaS product. There are no other users. The SLA is just: does this make my day better or not? But the discipline I've applied is the same as I would for a client system — proper orchestration logic, defined agent interfaces, a clear separation between the agents that think and the agents that act. Nova talks to my projects. It knows about AIREP. It can look things up, draft things, remind me of context I've forgotten. That only works if the plumbing is solid.</p><p>The thing I keep coming back to is that this is infrastructure, not tooling. There's a meaningful difference. A tool is something you pick up when you need it. Infrastructure is something that changes what's possible in the background, whether you're thinking about it or not. When I'm working on Find a Sign's marketplace model, Nova has context. When I'm making an architecture decision for AIREP, Nova has context. That ambient awareness is what I was building toward — not a smarter autocomplete, but a system that compounds knowledge across everything I'm working on.</p><p>One of the current goals I'm working toward is a self-improvement loop — agents that can review and refactor Nova's own code. That's not live yet, and I'm not going to pretend it is. But the production orchestration phase is what makes that goal reachable. You can't build a self-improvement loop on a prototype. You need a running system with real structure, real interfaces, and enough operational history that an agent reviewing the code has something meaningful to work with. Getting to production was the prerequisite.</p><p>The honest version of where things stand: Nova works. It's genuinely useful day-to-day. It's also incomplete in ways I see clearly — the self-improvement loop, better cross-project memory, tighter integration with AIREP's data. Those are next. But "works and incomplete" is the right state to be in. A system that's complete but not running is just a design document.</p><p>If you're building something similar — a personal AI system, an agent-based workflow, anything that involves coordinating multiple AI components — the advice I'd give is this: start with the orchestration layer before you build the agents. Know how tasks will be routed, how context will be passed, how failures will be handled. The individual agents are the easy part. The connective tissue is where the real work is, and it's much harder to retrofit than to design upfront.</p><p>Nova is the backend AI service visible on this site. If you're curious about the architecture in more detail, I'll keep writing about it here as it evolves.</p>