← Back to blog

Nova Hits Production Orchestration: 34 Agents, One Conductor

· 3 min read

Nova, my personal AI system, just crossed from a pile of capable agents into actual production orchestration. Here's what that shift actually means and why coordination — not capability — was the hard part.

<p>Nova just moved into production orchestration. That's a milestone I've been circling for a while, and it's worth writing down what it actually means — because the word "orchestration" gets thrown around a lot and usually means very little.</p><p>Nova is my personal AI system. Right now it's 34 agents spread across general chat, pipeline execution, ERP specialists, and language tasks. For a long time that number was the headline: more agents, more coverage. But a pile of capable agents isn't a system. It's a toolbox. The interesting and genuinely hard part was never adding the 35th agent — it was getting the existing ones to hand work to each other reliably, with the right context, and to know when to stop.</p><p>That's the shift production orchestration represents. Up to now, a lot of Nova's behaviour was me being the conductor: I'd decide which agent ran, feed it context, take its output, and route it onward. Useful, but it doesn't scale and it means the system is only as good as my attention on any given day. Orchestration means Nova decides the routing. A request comes in, gets classified, and the right agent — or sequence of agents — picks it up, with the context it needs already assembled.</p><p>The thing I underestimated was how much of this problem is about context, not capability. Each agent is, individually, fine. The failure mode in a multi-agent system isn't that an agent can't do its job — it's that it does its job with the wrong information, or hands off a half-formed result that the next agent can't use. Most of the work in getting to production was building the connective tissue: what each agent needs to know before it starts, what it's allowed to assume, and what it passes downstream. The agents are the easy part. The handoffs are where systems live or die.</p><p>There's a discipline this forced on me that I didn't expect. When you're the conductor, you can paper over sloppy boundaries with judgement. When the system orchestrates itself, every fuzzy boundary becomes a bug. An agent whose role is "sort of handles ERP questions, and also a bit of general stuff" is a liability. So orchestration pushed me to define roles sharply — this agent owns this, full stop — which made the whole thing more legible. That's a pattern I keep relearning: forcing automation onto something exposes exactly how vague your thinking was.</p><p>This connects to a broader bet I've been making about how I build. I don't treat AI as a feature I bolt onto products. I treat it as core leverage — the primary way I get work done across AIREP, Find a Sign, and my client work. Nova is the spine of that. The more it can route, sequence, and execute on its own, the more of my actual time goes to the decisions that need a human, and the less goes to being a glorified message router between models.</p><p>The next phase is the one I'm most interested in: a self-improvement loop. The idea is agents that review, refactor, and improve Nova's own code. That sounds ambitious, and it is, but orchestration is the prerequisite — you can't have agents safely working on the system until the system can reliably coordinate them. Production orchestration is the floor that makes the self-improvement ceiling reachable. Get the handoffs solid first; let the agents touch their own foundations second.</p><p>I'm deliberately not claiming Nova is "done" or that orchestration is fully solved. It works in production, which is a real and specific claim — it routes live requests and the agents cooperate without me holding their hands. But there's a long tail of edge cases, and there always will be. What's changed is the shape of the work. I've gone from being the thing that makes the agents useful to being the thing that decides what the agents should become. That's a better place to be standing, and it's the place I wanted to reach.</p><p>If there's a takeaway beyond my own project: in multi-agent systems, spend your effort on coordination and context, not headcount. The agents are commodity. The conductor is the product.</p>

Comments

No comments yet — be the first!

Leave a comment

Comments are held for moderation before appearing.