34 Agents and a Production System I Actually Use Every Day

<p>Nova hit a milestone recently that I keep having to explain to people: it's a personal AI system running 34 agents across general chat, pipeline execution, ERP specialists, language tasks, and more. The first question I usually get is "why do you need 34 agents?" It's a fair question, and the honest answer is that you don't start with 34 — you end up there.</p><p>The system started as a way to stop repeating myself. I work across several active projects simultaneously — AIREP, Find a Sign, Sweeper Parts, client websites, this portfolio — and context-switching is expensive. The early version of Nova was basically a smart notebook. Then it became a task router. Then I started wiring in specialised agents for specific domains, and now it's in what I'd call a production orchestration phase: it's not a prototype anymore, it's infrastructure.</p><p>What does "production" actually mean for a personal AI system? For me it means a few things. It means I rely on it for real decisions and real work, not just experiments. It means the architecture has to be stable enough that adding a new agent doesn't break existing pipelines. And it means I've accepted that this is a long-term engineering investment, not a side project I can abandon when something shinier comes along.</p><p>The 34 agents aren't arbitrary. They fall into clusters: general-purpose assistants, pipeline executors that chain tasks together, ERP-domain specialists that understand AIREP's data model and branch-scoped architecture, and language/writing agents that handle things like this blog post. Each cluster exists because I found myself doing the same class of work repeatedly with no good abstraction for it. The agents are the abstraction.</p><p>One thing I've learnt building this out is that the hard problem isn't the AI capability — it's the orchestration layer. Getting a single LLM to answer a question well is a solved problem for most practical purposes. Getting a system to correctly route a request, maintain context across multiple hops, pass structured data between agents without losing fidelity, and fail gracefully when something breaks — that's actual engineering. That's where most of the work lives.</p><p>The self-improvement loop goal is the next frontier I'm working toward. The idea is that Nova's agents should be able to review, refactor, and improve Nova's own code. It sounds recursive and slightly unhinged, but it's really just applying the same code-review and refactoring patterns I'd use as a developer — except the agent does the first pass. The human (me) still makes the call. I'm not interested in autonomous systems that run without oversight; I'm interested in systems that do the tedious groundwork so I can make better decisions faster.</p><p>This connects to a broader position I've been sharpening over the past year: AI is not a tool I use, it's a core competency I'm building. There's a meaningful difference. A tool is something you pick up when you need it. A competency is something that compounds — each thing you build makes the next thing cheaper and faster to build. Nova makes AIREP development faster. The patterns I develop in Nova inform how I think about AI integration in Find a Sign. The ERP specialists in Nova will eventually assist clients directly, not just me.</p><p>If you're a developer reading this and wondering whether it's worth investing this deeply in your own AI infrastructure — my honest take is: it depends on your time horizon. If you're optimising for the next three months, probably not. If you're optimising for the next three years, the compounding is real. I've spent significant time building Nova that I could have spent on billable work. But Nova now saves me more time per week than it cost to build, and that gap is widening.</p><p>The system is running. It's messy in places, over-engineered in others, and there are agents that probably deserve to be merged. But it's mine, I understand every part of it, and it's genuinely useful. That's the bar I hold everything to.</p>