Nova Hit Its Orchestration Phase, and 34 Agents Is When Coordination Becomes the Hard Part

<p>Nova just crossed into its production orchestration phase. For context, Nova is my personal multi-agent AI system — the thing that sits behind my workflow and, increasingly, behind keirantrace.com as a backend AI service. It now runs 34 agents spread across general chat, pipeline execution, ERP specialists, and language tooling. That number isn't a brag; it's the reason I'm writing this. Somewhere between five agents and thirty-four, the hard problem stops being any individual agent and becomes how they coordinate.</p><p>When you have a handful of agents, you can hold the whole system in your head. You know which one handles what, and the routing is basically a switch statement you could read in one sitting. At 34, that mental model breaks. You start getting the multi-agent equivalent of distributed-systems problems: who owns a task, what happens when one agent's output is another's input, how you avoid two agents quietly doing the same work, and how you keep the whole thing observable when a request fans out across several specialists.</p><p>So the orchestration phase isn't about adding more agents. It's about the layer that sits above them. The interesting questions are boring-sounding ones: how does a request get decomposed, how is work routed to the right specialist, and how do partial results get composed back into something coherent? Those are the questions that decide whether 34 agents feel like a team or feel like 34 people talking over each other in a meeting.</p><p>The thing that surprised me is how much this rhymes with ERP architecture. In AIREP — my multi-tenant Django ERP — the core discipline is branch-scoped data: each client is isolated at the branch level so nothing leaks across boundaries you didn't intend. Orchestrating agents is the same instinct in a different domain. Every agent needs a clear scope, a clear contract for what it accepts and returns, and a boundary that the orchestrator respects. The agents that cause trouble are exactly the ones with fuzzy scope — the ones that try to do a bit of everything. Tight boundaries make the system legible. That's true whether the boundary is around a tenant's data or around an agent's responsibility.</p><p>The next milestone I care about is the self-improvement loop: agents that can autonomously review, refactor, and improve Nova's own code. I'm deliberately not rushing that. A self-modifying system is only safe to the degree it's observable, and observability is precisely what the orchestration layer is meant to provide. You don't let a system edit itself until you can clearly see what it's doing and reliably stop it. So the unglamorous orchestration work now is the foundation for the more ambitious work later. Build the nervous system before you ask it to operate on itself.</p><p>There's a broader reason I'm investing here. I treat AI as the primary leverage point in how I build software and run the business — not as a feature bolted onto products, but as a core competency. AIREP, Find a Sign, and Nova are all meant to use AI as a genuine differentiator rather than a marketing line. Nova is where I prove that to myself first. If I can't make a 34-agent system coordinate cleanly for my own work, I've no business selling AI capability inside a client's ERP.</p><p>If there's a lesson from this stretch, it's that the difficulty in multi-agent systems migrates. Early on, the agents are the hard part — getting any single one to do a useful job. Later, the agents are the easy part and the orchestration is where everything either holds together or quietly falls apart. The shift caught me a little off guard, which is usually a sign you're working on the right problem.</p><p>I'll write more once the self-improvement loop is actually running rather than planned. For now, the honest update is that Nova is in production, coordinating real work across real agents, and the boring infrastructure is the thing making the interesting stuff possible.</p>