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

<p>At some point in the last few months, Nova stopped being an experiment and started being something I depend on. It's hard to pinpoint exactly when that shift happened, but there's a clear marker now: 34 agents in production, spanning general conversation, pipeline execution, ERP specialists, and language tasks. That's not a demo. That's a system.</p><p>I want to be honest about what "34 agents in production" actually means, because there's a lot of noise in the AI space about agent counts as if they're a score. The number isn't the point. What matters is that each agent has a defined role, a reason to exist, and a failure mode I understand. Some of them are thin wrappers around specific tasks. Others carry real orchestration logic — routing work, managing state, deciding what gets handed off and when. The architecture has teeth.</p><p>Nova is my personal AI system, and it runs as the backend service behind keirantrace.com — visible to visitors through the site itself, and used daily by me across every active project. AIREP, Find a Sign, Sweeper Parts, client websites — Nova touches all of it in some form. It handles queries, drafts content, assists with code, and increasingly takes on coordination work that used to live in my head or in scattered notes.</p><p>The shift I'm thinking about a lot right now is the difference between AI as a tool and AI as infrastructure. When it's a tool, you reach for it when you need it. You open a tab, type a prompt, get an output, close the tab. The value is real but it's episodic. Infrastructure is different — it's always running, it has state, it's part of the system rather than adjacent to it. A broken tool is inconvenient. Broken infrastructure stops work.</p><p>Nova is starting to feel like infrastructure. That's exciting and sobering at the same time. Exciting because the leverage is real — having 34 specialised agents available as callable components changes what's tractable for a solo developer running multiple projects simultaneously. Sobering because infrastructure demands a level of care that tools don't. You have to think about observability, about what happens when an agent behaves unexpectedly, about how to evolve the system without breaking what already works.</p><p>One of the goals I've been pushing toward is a Nova self-improvement loop — agents that can autonomously review, refactor, and improve Nova's own code. That goal made sense as an aspiration when Nova had ten agents. With 34 in production, it's becoming a practical necessity. The codebase is large enough that human-only maintenance doesn't scale the way I'd like. The agents need to carry some of the load of their own upkeep.</p><p>What I've learned building to this point: specialisation matters more than I expected. Early on I leaned toward general-purpose agents that could be prompted into different roles. That works at small scale. As the system grows, agents with tightly scoped responsibilities are dramatically easier to reason about, debug, and improve. The ERP specialists in Nova exist because AIREP has domain-specific logic that doesn't belong in a general-purpose assistant. Keeping that separation clean has paid off repeatedly.</p><p>I've also learned that orchestration is the hard part. Individual agents are relatively easy to build and test in isolation. Getting them to work together reliably — passing context cleanly, handling failures gracefully, avoiding loops or stalls — that's where the real engineering lives. It's closer to distributed systems design than it is to prompt engineering, and I think a lot of the "agent framework" tooling in the ecosystem underestimates that.</p><p>The production orchestration phase, as I'm calling it, is about stabilising what's there and extending the system's ability to coordinate across projects. Less about adding new agents for the sake of it, more about making the existing ones work together well enough that I can trust the outputs without babysitting every run. That's the bar. We're not there yet, but we're closer than we were six months ago, and the trajectory is clear.</p><p>If you're building agent systems yourself, my honest take: don't chase agent count. Chase clarity of role and reliability of handoff. The rest follows from there.</p>