AI as Leverage, Not Tooling: What Running 34 Agents in Production Actually Means

<p>There's a framing problem with how most developers talk about AI. They call it a tool. A productivity booster. A better autocomplete. That framing isn't wrong exactly, but it undersells what's actually available right now — and it leads people to integrate AI in the most shallow way possible: a chat window bolted onto the side of something they already built.</p><p>I've been thinking about this differently for a while, and Nova — my personal multi-agent AI system — is the concrete expression of that thinking. As of now, Nova is running in production orchestration across 34 agents. That covers general chat, pipeline execution, ERP specialists, language tasks, and more. It's not a demo. It's the backend AI layer I use every day, across every project I'm working on.</p><p>When I say "production orchestration phase," I mean agents are being coordinated to handle real workflows — not just answering questions, but executing multi-step tasks, passing context between agents, and operating with a degree of autonomy that would have been genuinely difficult to build even eighteen months ago. The infrastructure to do this at a useful level now exists, and the gap between "experimenting with AI" and "running AI as core infrastructure" is mostly a question of whether you're willing to commit to the architecture.</p><p>The distinction I keep coming back to is leverage versus tooling. A tool is something you pick up and put down. Leverage is structural — it changes the ratio of output to effort at a systemic level. If you're using AI to write a function here and there, that's a tool. If you're building systems where AI agents handle entire categories of work autonomously, that's leverage. The compounding effect is completely different.</p><p>For me, this shows up across three main projects right now. AIREP, my Django-based ERP system, is being built with AI integration as a first-class concern — not an afterthought. Find a Sign, an Australian signage marketplace, is designed around a customer-first discovery model where AI can do real work for buyers rather than just serving supplier ranking games. And Nova itself is on a trajectory toward a self-improvement loop: agents that can review, refactor, and improve Nova's own codebase. That last one is still a goal, not a current reality — but the architecture is being built with it in mind.</p><p>I want to be honest about what 34 agents in production actually looks like day-to-day. It's not magic. There are coordination problems. Context management is genuinely hard. Agents that work perfectly in isolation behave unexpectedly when they're chained. Debugging an agentic pipeline is a different skill to debugging a regular application — the failure modes are subtler and the state is harder to inspect. I've spent real time on these problems, and they're not fully solved.</p><p>But here's the thing: working through those problems is exactly where the advantage accumulates. Every developer who treats AI as a bolt-on is skipping the hard part. The hard part is the architecture — deciding how agents communicate, how context is preserved, how failures are handled, how you maintain oversight without micromanaging every step. Getting that right takes time and produces something that's genuinely difficult to replicate quickly.</p><p>I'm not making a prediction about the industry here. I'm describing what I'm actually building and why. The bet is simple: AI as a compound, structural advantage is worth the investment in getting the architecture right — and that investment pays off across every project it touches, not just one. That's why Nova exists as its own system rather than a feature inside something else. It's infrastructure, not a feature.</p><p>If you're a developer still treating AI as a productivity tool, I'd push you to ask a harder question: what would it look like to build AI into the structure of how you work, not just the surface? The answer is probably more architectural work than you want to do. But that's usually where the real leverage is.</p>