34 Agents and Counting: What Running Nova in Production Actually Looks Like

<p>Nova is in production. Not "production" in the sense that I demoed it to someone and it mostly worked — production in the sense that it's running, it's handling real tasks across multiple projects, and I depend on it daily. That milestone snuck up on me, so it's worth stopping to write down what it actually looks like.</p><p>Right now Nova has 34 agents. They span a few broad categories: general chat and reasoning, pipeline execution, ERP specialists for AIREP, and language tasks like summarisation and drafting. Each agent has a defined scope. None of them try to do everything. That constraint is intentional — the moment an agent becomes a generalist, it becomes unreliable in ways that are hard to debug. Narrow scope means predictable failure modes.</p><p>The orchestration layer is where most of the interesting engineering lives. Routing the right task to the right agent, managing context across a pipeline, knowing when to hand off versus when to loop — these are genuinely hard problems. They're not hard in a flashy algorithmic sense. They're hard in the way distributed systems are hard: the failure cases are subtle, they compound, and they usually surface at the worst possible moment. I've been working through them one at a time.</p><p>What I've found is that the value of a multi-agent system isn't in any single agent being impressive. It's in the compounding. An agent that drafts, another that reviews, another that reformats for a specific output target — individually they're fine. Together they replace a class of work that used to take me an hour and now takes seconds. That compounding is what I mean when I say AI is leverage, not tooling. A tool does one thing when you pick it up. Leverage changes what you're capable of at a structural level.</p><p>The self-improvement loop is the next thing I want to build. The idea is straightforward in concept: agents that can review Nova's own code, flag patterns that are getting brittle, and propose refactors. In practice it requires a level of trust in automated output that I'm not ready to grant blindly. So the first version will be advisory — agents surface issues, I make the call. That's the right boundary for now. Fully autonomous refactoring on a system I depend on would be reckless, and I'd rather be honest about that than oversell what the system can do.</p><p>Running this across multiple active projects — AIREP, Find a Sign, Sweeper Parts, client work — has also clarified something about how I want to structure my time. The projects that benefit most from Nova are the ones where the work is repetitive but context-sensitive. Generating ERP-specific documentation, summarising a thread of decisions on a project, drafting copy that needs to match a specific tone. Nova handles the first pass. I handle the judgment calls. That division isn't a compromise — it's actually the right split of cognitive labour.</p><p>There are rough edges. Context management across long pipelines is still messier than I'd like. Some agents are more reliable than others and I haven't fully characterised why. The orchestration logic has grown organically and needs a proper audit. None of that is surprising for a system at this stage, but I'm naming it because I think there's a tendency in AI writing to skip straight from "I built a thing" to "it's transformative" without spending any time in the middle, where the thing is real and imperfect and you're figuring out where the limits actually are.</p><p>That middle is where I am. 34 agents, production orchestration, genuine daily utility — and a clear list of things that still need work. That feels like the honest version of a milestone.</p>