Nova Has 34 Agents in Production — Here's What That Actually Means

<p>A few months ago Nova was a handful of Python scripts wired together with good intentions. Today it's a multi-agent system running 34 specialised agents across general chat, pipeline execution, ERP specialists, language tasks, and a growing set of domain-specific workers. I've started calling this the production orchestration phase, and it's worth being precise about what that means — because "AI agents in production" is one of those phrases that can mean almost anything.</p><p>In Nova's case, production orchestration means that agents are no longer just available — they're being routed to, composed, and trusted to carry work across session boundaries. When I ask Nova to do something non-trivial, the system decides which agent or chain of agents handles it, passes context between them, and returns a result I can actually act on. That routing logic is the hard part. Getting one agent to do a thing is easy. Getting the right agent to pick up the right task with the right context, without me babysitting the handoff, is a different problem entirely.</p><p>The 34 agents aren't 34 copies of the same thing. They're genuinely specialised. Some handle general conversation and memory. Others are scoped tightly to AIREP — understanding branch-level data isolation, Django model structure, ERP business logic. A separate cluster handles language tasks: drafting, summarising, rewriting. The pipeline execution agents are the ones I rely on most heavily right now — they take a defined task, break it into steps, execute, and report back. That's where most of the real leverage sits.</p><p>What I've learned building this is that specialisation is worth the overhead. Early on I tried to keep agents general — one smart agent that could handle most things. That falls apart quickly. A general agent optimises for breadth and ends up mediocre at everything that requires depth. The ERP agents I've built know things a general agent would have to rediscover every time: what branch-scoping means, how multi-tenant data should be isolated, what a typical AIREP workflow looks like. That embedded domain knowledge is what makes them actually useful rather than just plausible.</p><p>The other thing production has clarified is the importance of memory architecture. Agents that can't remember prior context are expensive to use — you're constantly re-explaining things. Nova's memory system now spans core profile data (injected into every session), archival notes (searchable on demand), and episodic memory (recent conversation history). Getting those tiers right has been as important as the agents themselves. An agent is only as useful as the context it can access.</p><p>I want to be honest about where this is still rough. Orchestration at this scale introduces failure modes that are harder to debug than a single-agent system. When something goes wrong, it's not always obvious which agent in the chain made the bad call. Logging and observability are things I'm actively improving. The system is useful enough that I'm relying on it daily, but it's not a finished product — it's a working foundation that I'm extending in parallel with using it.</p><p>The longer-term goal is a self-improvement loop: agents that can review Nova's own code, identify weak spots, and propose or apply fixes. That's still ahead of where I am today. But reaching production orchestration with 34 agents is a real waypoint on that path. The architecture is stable enough to build on, the agents are specialised enough to be genuinely useful, and the memory system is coherent enough that context survives across sessions.</p><p>If you're building something similar — a personal AI system, an internal agent framework, anything beyond a single-model chatbot — the lesson I'd pass on is this: invest in the plumbing before you invest in the agents. Routing, memory, and observability are what separate a system that works from a collection of demos. The agents are the visible part, but the infrastructure underneath is what makes them worth running.</p>