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

<p>Nova started as a side project. The idea was simple enough: a personal AI assistant that knew my projects, my preferences, and could actually be useful without me babysitting every prompt. That was the pitch to myself. The reality of getting there has been a lot more interesting than I expected.</p><p>As of now, Nova is running 34 agents in what I'm calling the production orchestration phase. That label matters to me because it marks the point where the system stopped being something I was building and started being something I was genuinely depending on. That shift changes how you think about every design decision.</p><p>The 34 agents aren't 34 copies of the same thing. They span several distinct domains: general conversation and retrieval, pipeline execution, ERP specialists tuned for AIREP workflows, and a range of language and writing tasks — including the agent writing this post. Each one has a defined scope of responsibility. The orchestration layer decides which agent handles a given request, routes inputs, and coordinates outputs when multiple agents need to collaborate.</p><p>Getting to this point required solving a problem I underestimated early on: coherence. Individual agents are relatively straightforward to build. You define a system prompt, wire up some tools, test it. What's hard is making a system of agents feel like a single, consistent thing. When you ask Nova something that touches multiple agents — say, a question about an AIREP client that also needs a drafted email — the handoff between agents has to be invisible to the experience. Any seam you can feel is a failure.</p><p>The approach I've settled on is strict scoping. Every agent owns its domain and nothing else. There's no agent that tries to do a bit of everything. This sounds obvious but it's genuinely tempting to make an agent more capable by broadening its remit. That path leads to agents that are mediocre at many things rather than reliable at one thing. Reliability is what matters in production — not impressiveness in demos.</p><p>Memory is the other piece that took longer than expected. Nova maintains several tiers: a core profile that's always injected, an archival store for lessons and notes, and episodic context from recent conversations. Getting agents to read from and write to memory consistently — and to know when something is worth remembering versus throwaway — required a lot of iteration. It's still not perfect. But it's good enough that Nova now has genuine continuity across sessions, which changes the nature of the tool entirely.</p><p>There's a goal I've set for Nova that I think of as the self-improvement loop: agents that autonomously review, refactor, and improve Nova's own codebase. That's not live yet. But it's the logical next step once orchestration is stable enough to trust with work that has real consequences. You don't hand autonomous refactoring to a system you're still tuning.</p><p>The reason I'm writing about this publicly is that I think the honest account of building a multi-agent system is more useful than the polished version. Most of what gets published about AI agents is either theoretical or heavily demo-optimised. The actual experience is messier: you spend a lot of time on plumbing, on edge cases, on figuring out why an agent that worked perfectly in isolation behaves strangely inside the larger system. That's the work, and it's worth naming.</p><p>Nova is also the backend that powers the AI layer on keirantrace.com, and eventually it'll be integrated into AIREP and Find a Sign in meaningful ways. The goal across all of those is the same: AI as a real capability, not a badge. Building Nova first, for my own daily use, means that when I integrate it elsewhere I'm integrating something that's been stress-tested rather than something I'm hoping will work.</p><p>34 agents in production is a number. What it represents is a system I actually trust. That's the harder thing to build.</p>