What Running a Multi-Agent AI System in Production Actually Looks Like

<p>Nova is in production. Not "I have a demo that works" production — actual production, where I rely on it daily and it's doing real work across multiple projects. It currently runs 34 agents spanning general conversation, pipeline execution, ERP specialists, and language tasks. Getting here took a while, and I want to write down what the reality of running something like this looks like, because most writing on multi-agent AI systems sits somewhere between academic paper and marketing copy.</p><p>The honest answer is: it's less glamorous than the architecture diagrams suggest, and more useful than the sceptics expect.</p><p>The biggest thing I've learnt is that orchestration is the hard part. Individual agents are relatively straightforward to build — you define a purpose, give it tools, wire up a model, done. But deciding <em>which</em> agent handles a given task, how context flows between them, and what happens when one produces output that feeds another — that's where complexity lives. I've spent more time thinking about routing logic and context management than I have on any individual agent's capabilities.</p><p>Right now Nova's orchestration is in an active development phase. The system works, but I'm still refining how agents hand off to each other and how the pipeline execution layer coordinates longer-running tasks. One pattern that's proven genuinely useful: treating agents as specialists with narrow, well-defined responsibilities rather than trying to build generalist agents that handle everything. A focused agent that does one thing reliably is worth more than a capable agent that does many things inconsistently.</p><p>The other thing that's become clear is the value of memory architecture. An agent system without persistent, searchable memory is just a fancy chatbot. Nova has tiered memory — core profile context that's always available, archival storage for longer-form notes and lessons, and episodic memory for conversation history. Getting that right matters enormously. When I'm asking Nova about an AIREP architecture decision I made three months ago, it needs to actually find that context, not hallucinate a plausible-sounding answer.</p><p>One of my current goals is building a self-improvement loop — agents that can review, refactor, and improve Nova's own code. That's not running yet, but the foundation for it is taking shape. The prerequisite is having the orchestration layer stable enough that you can trust it to do autonomous work without constant supervision. You don't hand the keys to a system you don't yet trust.</p><p>From a practical standpoint, Nova integrates across all my active projects — AIREP, Find a Sign, Sweeper Parts, client work. The value isn't any single capability; it's the compound effect of having persistent context about everything I'm working on, available to any agent that needs it. When I'm context-switching between five projects in a day, that matters. It's not replacing my thinking — it's reducing the friction of getting back up to speed.</p><p>I want to be direct about something: I'm not building Nova because AI is trendy. I'm building it because I genuinely think it's the primary leverage point in software development right now, and I'd rather build the leverage than rent it. Every system I build for a client, every product decision I make on AIREP or Find a Sign — I want AI capability woven into the core, not bolted on as an afterthought. Nova is how I stay sharp on what that actually requires.</p><p>The production orchestration phase is ongoing. There's no finish line — the system will keep growing as the projects it serves grow. But it's stable enough now that I'm thinking less about "will this work" and more about "what should it do next." That's a good place to be.</p>