Why I Don't Want One AI, I Want Thirty-Four

<p>People ask me what Nova actually is and I usually give the short answer: a personal AI system. The longer answer is that it's not one thing at all. It's currently 34 agents, split across general chat, pipeline execution, ERP specialists, and language tasks. That number is going to keep changing, but the architecture behind it — many narrow agents instead of one broad one — is the part I actually care about.</p><p>The instinct when you start building with AI is to reach for the biggest, most general model and point it at everything. That works fine for a demo. It falls apart the moment you're running AIREP, Find a Sign, Sweeper Parts, and a handful of client sites at the same time, each with its own data shape, its own domain rules, its own failure modes. A general-purpose agent answering ERP questions one minute and signage supplier queries the next doesn't get better at either — it gets mediocre at both, and worse, it doesn't know what it doesn't know.</p><p>Specialist agents fix that, but not for the reason people assume. It's not really about the model being smarter when scoped narrowly. It's about being able to reason about the system as an engineer. When an ERP specialist agent gives a wrong answer about branch-scoped data, I know exactly where to look — the agent's context, its prompt, its access to the AIREP schema. When a single monolithic assistant gets something wrong, I'm debugging a black box with no seams. Specialisation gives you seams. Seams are where you fix things.</p><p>This matters more as I push toward treating AI as a core business competency rather than a feature bolted onto existing products. AIREP, Nova, and Find a Sign all need AI to be a genuine differentiator, not a chatbot widget in the corner. That only works if the AI layer is architected with the same rigour as the rest of the stack — multi-tenant isolation, clear boundaries, agents that know their lane. A sprawling do-everything model doesn't respect those boundaries. A swarm of scoped agents can.</p><p>There's a business reason for this too, separate from the engineering one. Find a Sign is built on the idea that customers should get transparent, supplier-first discovery — no pay-to-rank, no artificial incentive to favour one supplier over another. I want the AI side of my projects to hold the same standard internally: no single agent quietly becoming the one that decides everything because it's the one that happens to be loaded by default. Specialisation is a kind of honesty constraint. Each agent does its job and nothing more, which means nobody — including me — can accidentally let one piece of the system accrue power it shouldn't have.</p><p>The next piece I'm working toward is a self-improvement loop — agents that can review and refactor Nova's own code rather than me manually tuning 34 separate configurations by hand. That's a harder problem than it sounds, because self-review only works if the reviewing agent has a narrower, more trustworthy scope than the thing it's reviewing. You can't let a general-purpose agent mark its own homework. You need something closer to a code-review specialist with a fixed rubric, looking at a fixed slice of the system, producing a diff someone (eventually an agent, for now me) can actually evaluate.</p><p>None of this is me arguing against big general models — they're useful, and some of the specialist agents in Nova are thin wrappers around exactly those models with tighter prompts and narrower context. The point isn't model size, it's system design. A personal AI system that's going to run real businesses needs the same discipline as a multi-tenant ERP: isolated scope, clear ownership, auditable boundaries. Thirty-four agents sounds like a lot until you realise the alternative is one agent pretending it can do thirty-four jobs at once.</p>