Nova Is in Production: What 34 Agents Actually Looks Like

<p>Nova is in production. Not "I ran it once and it worked" production — actual orchestration-phase production, where the system is handling real tasks across real projects and I'm trusting it with things that matter. That's a different line to cross than getting a prototype to demo well.</p><p>The system currently runs 34 agents. I know that number sounds like something you'd put in a press release, so let me be honest about what it actually means. The agents aren't 34 instances of the same chat loop with different system prompts. They break down across distinct functional categories: general conversation and reasoning, pipeline execution, ERP specialists that understand AIREP's data model, and language-focused tasks like writing and summarisation. Each category exists because I hit a real problem that a general-purpose agent handled badly. The specialisation is earned, not architectural decoration.</p><p>Getting to orchestration phase meant solving a different class of problem than building the first agent. A single agent is mostly a prompt engineering and API integration challenge. An orchestrated system is a software architecture challenge. You have to answer questions like: how does one agent hand off to another without losing context? How do you avoid an orchestrator that just becomes a bottleneck — a single-threaded dispatcher wearing a fancy hat? How do you handle failures gracefully when the failing component is itself an AI system with non-deterministic outputs?</p><p>I don't have perfect answers to all of those yet. But the production phase is where you find out which questions matter most. Running Nova against real work — client project tasks, AIREP development, writing, research — surfaces failure modes that no amount of local testing catches. The gap between "works in isolation" and "works in a live workflow" is where most of the interesting engineering lives.</p><p>The thing I keep coming back to is that the value of a system like this isn't in any individual agent. It's in the compound effect over time. When the system helps me move faster on AIREP architecture decisions, that time goes somewhere — usually into thinking more carefully about the next problem, or shipping something I'd otherwise have deferred. The leverage stacks. That's the actual pitch for building this, not the agent count.</p><p>I've been deliberate about keeping Nova personal — it's my AI system, built around how I work. It's not a product, at least not yet. That means I can make opinionated choices that a commercial system couldn't: I can give it deep context about AIREP's multi-tenant branch-scoped data model, about how Find a Sign's marketplace logic works, about my specific values around transparency and not manipulating customers. A generic assistant doesn't know any of that, and getting it to behave consistently with those values requires either constant prompting or a system that's actually been built with them in mind.</p><p>The next phase I'm focused on is the self-improvement loop — agents that can review and refactor Nova's own code. That's a harder problem than it sounds. It's not just "have the AI write some code". It requires the system to have a stable enough representation of its own architecture to reason about changes without introducing regressions. I haven't solved that yet. But having 34 agents in production orchestration gives me a real target to improve against, which is a better position than theorising about it.</p><p>If you're building something similar — or thinking about it — the most honest advice I can give is: get something into production as fast as you can, even if it's small. The architectural problems you'll face in a live system are different from the ones you'll anticipate on a whiteboard, and the faster you're solving real problems, the faster you learn what actually matters.</p>