34 Agents in Production: What Running a Real Multi-Agent System Actually Looks Like

<p>I've been building Nova for a while now, and somewhere in the last few months it crossed a threshold I didn't have a clean name for until recently: it's in production. Not production in the sense that I'm serving external users — but production in the sense that I'm genuinely depending on it. It routes my own work. It's the backend for this site. It handles real tasks I would otherwise do manually. That's a different bar than "it runs on my machine."</p><p>Right now Nova runs 34 agents. That number sounds arbitrary, but it reflects how the system actually grew: general chat agents, pipeline execution agents, ERP specialists for AIREP, language and summarisation tasks, and a few that exist purely to coordinate other agents. The architecture is multi-agent by necessity, not by design philosophy. Each agent does one thing well. The orchestration layer decides which one to use and when.</p><p>What nobody really talks about when they describe multi-agent systems is how much the interesting problems are operational, not conceptual. The hard part isn't building an agent that can do a task — it's knowing when to invoke it, what to pass it, what to do when it returns something unexpected, and how to recover gracefully rather than silently failing. These are boring infrastructure problems. They're also the problems that determine whether the system is actually useful or just impressive in demos.</p><p>The self-improvement loop is the next thing I'm working toward. The idea is that Nova's agents should be able to review, refactor, and improve Nova's own code — under human oversight, at least initially. This isn't science fiction; it's a natural extension of what I'm already doing. I already use Nova to help me write and review code across my other projects. Pointing that capability at itself is a logical next step, and it's one of the things that makes the multi-agent architecture valuable in a way a single model never could be. You can have one agent write, another review, another test, without any one of them needing to hold the full context of the system in a single context window.</p><p>What I've learnt building this is that AI as a compound advantage is real, but it doesn't come from using AI tools occasionally. It comes from integrating AI deeply enough into your workflow that it changes the economics of what you can build alone. I run multiple active projects — AIREP, Find a Sign, Sweeper Parts, client work, this site — and the honest answer to how that's possible is Nova. Not because it does everything, but because it reduces the cost of context-switching, record-keeping, and the kind of low-level cognitive overhead that accumulates and slows you down.</p><p>The next goal I'm thinking about seriously is a unified project dashboard — a single workspace that shows work-in-progress, open issues, and recent changes across all my active projects. Right now that information is scattered, and even with Nova's help there's friction in getting a coherent picture of where everything stands. A dashboard doesn't sound glamorous, but it's the kind of thing that makes compounding possible: if you can see the whole board at once, you make better decisions about where to put time.</p><p>I don't have a tidy conclusion here. Nova is working, it's useful, and it's still being built. That's the honest state of it. The gap between "a multi-agent system that runs" and "a multi-agent system that reliably improves itself" is significant, and I'm not going to pretend I've closed it. But the first gap — between prototype and something I genuinely depend on — is closed. That felt worth writing down.</p>