Nova Is in Production: What Running 34 AI Agents Actually Looks Like

<p>A few months ago, Nova was a sketch — a personal AI system I wanted to build because nothing off the shelf fit how I actually work. This week it crossed a threshold I've been pushing toward: production orchestration. Thirty-four agents, running across general chat, pipeline execution, ERP specialists, language tasks, and more. It's not a demo. It's live infrastructure that I use every day.</p><p>I want to be honest about what that means, because "34 AI agents" sounds like a marketing slide. In practice, it means I have a coordinator that routes incoming tasks to specialist agents based on context, and those agents have discrete responsibilities — some handle AIREP-related queries, some work on language and writing tasks, some run structured pipelines with defined inputs and outputs. They don't all run simultaneously. The system is closer to a staffing model than a hive mind: the right agent gets engaged for the right task.</p><p>Building this forced me to make decisions I hadn't fully thought through when I started. The biggest one was how to handle state and memory across agents. Each agent needs enough context to be useful, but you can't just dump everything into every prompt — that gets expensive and slow, and it dilutes focus. I ended up with a tiered memory model: core profile data that's always injected, archival notes retrieved on demand, and episodic conversation history that gets searched semantically. It's not perfect, but it works well enough that I rarely have to re-explain context to Nova mid-task.</p><p>The other thing that became clear quickly: agent boundaries matter more than agent count. Early on I was tempted to just build one very capable general agent and call it done. The problem is that general agents compromise. They're okay at everything and great at nothing. When I'm working on AIREP's multi-tenant data model, I want an agent that understands Django ORM patterns, branch-scoped isolation, and PostgreSQL schema design — not one that also has to hold opinions about email tone and Python packaging simultaneously. Specialisation makes the outputs noticeably better.</p><p>One of the things I'm building toward is a self-improvement loop — agents that can review and refactor Nova's own code. That's still in progress. But even now, Nova handles a meaningful chunk of the cognitive overhead of running multiple projects in parallel. I'm juggling AIREP, Find a Sign, Sweeper Parts, client sites, and this blog. Without a system that can hold context across all of those and surface the right information at the right time, I'd be spending hours a week just on orientation — remembering where I left off, what the open questions are, what I decided last week and why.</p><p>There's a broader point here about AI as leverage. I don't think of Nova as a chatbot I happen to own. It's closer to infrastructure — the same way a well-designed database schema or a solid CI pipeline is infrastructure. It compounds. Every agent I add, every memory tier I refine, every pipeline I define makes the system more useful across everything I build. That's the bet: that AI as a core operational competency, built into the fabric of how I work rather than bolted on as a feature, is the right long-term play.</p><p>The honest caveat is that this took real engineering time. It wasn't plug-and-play. I had to think carefully about routing logic, about what belongs in core memory versus archival storage, about how to make pipelines composable without making them brittle. If you're thinking about building something similar, my advice is to start smaller than you think you need to and be ruthless about agent scope. One agent doing one thing well is more useful than five agents with overlapping responsibilities and unclear ownership.</p><p>Nova being in production doesn't mean it's finished. The self-improvement loop is next. A unified dashboard across all my active projects is on the list. But getting to this point — real agents, real tasks, real daily use — is a meaningful milestone. It's the difference between a side project and a system. I'll keep writing about what I learn as it evolves.</p>