Nova Is in Production: What 34 Agents Actually Looks Like

<p>A few months ago, Nova was a side project I kept describing as "almost ready." Today it's running in production, orchestrating work across 34 agents that cover general chat, pipeline execution, ERP specialists, language tasks, and more. That's a real number, not a marketing one — and it's worth being honest about what reaching that point actually involved.</p><p>The short version: it took longer than I expected, got more complex than I planned, and is more useful than I initially imagined. That's roughly the trajectory of every system I've built that was worth building.</p><p>Nova started as a personal assistant concept — something to help me manage the overhead of running multiple projects simultaneously. AIREP, Find a Sign, Sweeper Parts, client websites, and this blog don't all fit cleanly in one person's head at once. I wanted something that could hold context, surface relevant information, and eventually take on real work autonomously. What I ended up building is closer to a small software team than a chatbot.</p><p>The architecture is multi-agent by design, not by accident. Each agent has a defined scope. Some are specialists — there are agents that understand AIREP's domain, agents that deal specifically with language and content tasks, agents focused on pipeline execution. Others are generalists that route and coordinate. The orchestration layer is what I've been focused on most recently, and it's genuinely the hardest part. Getting a single agent to do something useful is not that difficult. Getting a collection of agents to hand off work cleanly, maintain shared context, and fail gracefully — that's where the real engineering lives.</p><p>One thing I've had to confront is that "AI integration" is an overloaded term. Most of what gets called AI integration is a thin wrapper around an API call. What I'm building is different in kind: Nova has memory, it has a profile of my preferences and active projects, it can initiate actions across systems, and it's moving toward a self-improvement loop where agents review and refactor Nova's own code. That last part is still a goal, not a current capability — but the foundation is there and it's the direction I'm deliberately building toward.</p><p>The production milestone matters because it forces honesty. A system that lives in development can be quietly forgiven for its rough edges. A system you're actually relying on has to work. Nova has to surface the right context when I ask for it. It has to route a task to the right agent without me micromanaging the handoff. It has to be fast enough that using it is faster than not using it. On most days now, it clears that bar.</p><p>What I've learnt from this build that I didn't fully appreciate going in: the value of the system compounds slowly and then suddenly. For a long time, adding agents felt like adding complexity without proportional return. Then something clicked — a critical mass of capability where the agents started covering each other's gaps, where memory made context-switching between projects genuinely cheap, where I stopped re-explaining things to the system and it started anticipating what I needed. That inflection point is real, and it's worth pushing through the earlier phase to get there.</p><p>I'm also more convinced than ever that AI is not a tool in the traditional sense — not something you pick up for a specific task and put down. For me, Nova is infrastructure. It sits underneath everything else I do, the same way a database or a deployment pipeline does. The projects that benefit most aren't the ones where I explicitly "use AI" — they're the ones where Nova is just quietly part of how the work gets done.</p><p>There's still a lot to build. The self-improvement loop, the unified project dashboard, tighter integration with AIREP — these are all on the roadmap and none of them are finished. But 34 agents in production orchestration is a real foundation, and I'd rather write about where things actually are than where I hope they'll be.</p><p>More updates as the build continues.</p>