What Running 34 AI Agents in Production Actually Looks Like

Nova, my personal AI system, has reached production orchestration phase with 34 specialised agents. Here's what that milestone actually means — and why the hard part was never the AI.

Nova is my personal AI multi-agent system. As of now, it's in what I'm calling the production orchestration phase — meaning it's not a prototype, not a proof of concept, not a demo I spin up occasionally. It's running, it's handling real tasks, and it's integrated into how I work across multiple projects. There are currently 34 agents covering general chat, pipeline execution, ERP specialists, language tasks, and more.I want to be honest about what that means, because "multi-agent AI system" sounds more like a product announcement than a working reality. So let me describe what it actually is.Each agent in Nova is scoped to a domain. There are agents that know AIREP's architecture. There are agents that understand Find a Sign's marketplace model and the values behind it. There are language-focused agents, pipeline agents that handle chained task execution, and a blog writer agent that produced this post. The idea is that no single agent needs to be a generalist across everything — it just needs to be good at its slice. Orchestration is what ties them together.Getting to 34 agents wasn't the challenge. The challenge was discipline: figuring out what each agent should and shouldn't know, where the boundaries sit, and how to pass context between them without creating a system where everything depends on everything else. That's a software architecture problem, not an AI problem. The same instincts that apply to designing a clean Django service or a well-scoped PostgreSQL schema apply here. Separation of concerns doesn't stop mattering just because your modules are LLM-backed.The production orchestration phase specifically means I'm now focused on how agents coordinate — how a task handed to one agent can trigger another, how output from a pipeline agent feeds downstream work, and how the system holds context across a session without losing coherence. It's less about adding more agents and more about making the existing ones compose cleanly.One of my current goals is to build a Nova self-improvement loop: agents that autonomously review, refactor, and improve Nova's own code. That's not running yet, but the groundwork is there. The reason I think it's achievable isn't because AI is magic — it's because the codebase is disciplined enough that an agent with the right context can reason about it. Messy codebases don't get better with AI assistance; they get messier faster. That's a lesson worth learning before you build something like this.The other thing I want to push back on is the framing of AI as a productivity tool. I've written about this internally and I keep coming back to the same position: AI isn't a tool I reach for to go faster. It's a core competency I'm building into the architecture of everything I make — AIREP, Find a Sign, Nova itself. The difference matters. A tool you reach for is optional. A core competency shapes how you design from the start.Nova is the clearest expression of that. It exists because I wanted to understand what it actually takes to run AI in production — not demo it, not wrap an API call in a button, but integrate it into real workflows with real constraints. Thirty-four agents in and I can say: the AI part is increasingly the easy part. Memory management, context scoping, orchestration logic, agent boundaries — that's where the engineering work lives.I'll keep writing about this as it develops. Next up is the self-improvement loop, and I want to document that honestly — including what doesn't work.