Nova Hits Production Orchestration: 34 Agents, One Conductor

<p>Nova just moved into its production orchestration phase, and I want to write down what that actually means before the milestone blurs into the next one.</p><p>For a while now Nova has been a collection of capable parts: 34 agents spread across general chat, pipeline execution, ERP specialists, and language tooling. Each one does its job well in isolation. The interesting and genuinely hard work was never making any single agent smarter. It was making them work together without me babysitting the handoffs.</p><p>That distinction took me longer to internalise than I'd like to admit. When you build agents one at a time, the obvious next move always looks like "add another capability." But a pile of capable agents is not a system. It's a staffing problem. The bottleneck shifts from what each agent can do to who decides which agent does it, in what order, with what context, and what happens when one of them returns garbage. Orchestration is that decision layer, and it's where the real engineering lives.</p><p>Production orchestration, for me, means a few specific things. Requests now get routed to the right agent without me hand-wiring the path. Multi-step work flows through a pipeline where the output of one stage becomes the structured input of the next. Specialists — the ERP agents, the language tooling — get pulled in on demand rather than always sitting in the loop. And critically, it's reliable enough that I trust it with real work instead of treating every run as an experiment I need to supervise.</p><p>The thing nobody tells you about multi-agent systems is that context management is the whole game. An agent is only as good as what it knows at the moment it acts. Too little context and it confidently does the wrong thing. Too much and you blow your token budget shuttling irrelevant history between stages. A lot of the orchestration work was deciding what each agent needs to see, what it should forget, and where state gets persisted between steps. That's not glamorous, but it's the difference between a demo and a system you actually use.</p><p>I also had to get honest about failure handling. In a single-agent setup, a bad response is a bad response — you see it and retry. In a pipeline, a bad response at step two silently poisons steps three through six, and you don't notice until the end looks wrong. So orchestration isn't just routing; it's knowing when to stop, when to retry, and when to escalate to me. An agent that knows when to ask is worth more than one that always answers.</p><p>This connects to a broader bet I've been making across all my work — AIREP, Find a Sign, the client projects. I don't treat AI as a feature I bolt on. I treat it as the primary leverage point, the thing that compounds. Nova is where I prove that to myself first, because it's mine to break. Whatever orchestration patterns hold up here are the ones I'll trust enough to put in front of clients.</p><p>The next phase is the part I'm genuinely excited about: a self-improvement loop. Agents that review, refactor, and improve Nova's own code. There's an obvious circularity there that makes it both interesting and slightly dangerous — a system editing the system that edits it. I'll be keeping that on a very short leash, with human review on anything that touches the orchestration layer itself. Compound advantage is real, but so is compounding mistakes, and I'd rather be the one who notices the difference.</p><p>For now, I'm letting the production phase settle. There's a temptation after a milestone to immediately pile on more agents, more capabilities, more scope. I'm resisting it. The win this week wasn't doing more — it was getting 34 things I already had to actually cooperate. That's a better foundation than any single new feature would have been.</p>