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

<p>Nova is in production. Not "works on my machine" production, not a polished demo — actual production orchestration, handling real tasks across general chat, pipeline execution, ERP specialists, language tools, and more. As of right now it's running 34 agents. I want to write about what that actually feels like to build and operate, because most of what's published about multi-agent AI systems is either academic or a sales pitch.</p><p>The honest version: it's less like conducting an orchestra and more like being a plumber who also designed the pipes. When it works, it's genuinely impressive — a task comes in, the right agent picks it up, context flows through correctly, and something useful comes out the other side without me touching anything. When it doesn't work, you're reading logs trying to figure out whether the failure was in routing, context passing, tool execution, or the model itself. The debugging surface is large.</p><p>The architecture I landed on groups agents by domain. There are general-purpose chat agents, pipeline agents that execute multi-step workflows, ERP specialists that understand AIREP's data model and business logic, and language/text agents for writing and summarisation tasks. Each agent has a defined scope — they don't try to do everything, and that constraint is deliberate. The moment an agent becomes a generalist, it becomes harder to reason about what it should and shouldn't do. Narrow scope means clearer failure modes.</p><p>One thing I underestimated early on was the orchestration layer itself. It's easy to build ten agents that each work in isolation. The hard part is the system that decides which agent handles a given task, how context gets preserved across handoffs, and what happens when an agent returns a result that needs further processing by a different agent. That layer is where most of the real engineering lives — and it's also where most of the subtle bugs hide.</p><p>Context management is the other thing that doesn't get talked about enough. Each agent needs enough context to do its job, but you can't just pass everything everywhere — it's expensive, slow, and introduces noise. So you end up building something like a selective memory system: what does this agent actually need to know right now? For Nova, I've been building out tiered memory — profile-level facts that persist across conversations, archival notes that are retrievable by search, and episodic context for recent activity. Getting that layering right is an ongoing project, not a solved problem.</p><p>The goal I'm working toward now is a self-improvement loop — agents that can review Nova's own code and behaviour, identify weak spots, and propose or implement changes. That's not science fiction; the infrastructure is mostly there. What's missing is the evaluation layer: how does the system know whether a proposed change is actually better? That's the hard part, and I don't think there's a shortcut. You need real usage, real failure cases, and honest measurement.</p><p>What I've found most valuable about building Nova isn't any single capability — it's the compound effect. Each agent I add makes the system more useful across all my projects. Nova already has context about AIREP's architecture, Find a Sign's marketplace model, Sweeper Parts' domain, and my client work. When I'm making a decision on one project, I can query that context and get something genuinely useful back — not because the model is magic, but because the memory and agent design make the right information available at the right time.</p><p>That's the version of AI I'm building toward: not a chatbot bolted onto a product, but a system that has real, persistent knowledge of my work and can act on it. 34 agents is a milestone, but it's mostly a number. The real measure is whether the system makes me faster, surfaces things I'd otherwise miss, and compounds in usefulness over time. So far, the answer is yes — with caveats, and with a lot of work still to do.</p>