34 Agents in Production: What Building Nova Actually Taught Me

<p>A few months ago, Nova was a side project. A personal assistant with a handful of tools bolted together. Now it's a production orchestration system running 34 agents across general chat, pipeline execution, ERP specialists, language tasks, and more. It's the backend AI layer behind this website. It manages memory, drafts content, talks to my other systems, and increasingly runs unsupervised on tasks I used to handle manually.</p><p>I want to write honestly about what that process has actually been like — because most writing about multi-agent AI systems is either academic or promotional. This is neither.</p><p>The first thing I got wrong was thinking the hard part was the AI. It isn't. The models are capable. The hard part is the plumbing: routing, memory, context propagation, and knowing when an agent should act versus when it should ask. Those are engineering problems, not AI problems. Once I stopped treating agent design as a prompt-engineering exercise and started treating it as a software architecture problem, things clicked.</p><p>The second thing I got wrong was scope. I started with a general-purpose agent and kept adding capabilities to it. That works until it doesn't — and it stops working faster than you'd expect. A single agent trying to do everything becomes incoherent. Its context window fills with irrelevant history, its instructions become contradictory, and its outputs get mushy. Breaking Nova into specialised agents with clear, narrow responsibilities made each one dramatically more reliable. A pipeline execution agent that only knows about pipelines is better at pipelines than a general agent that also knows about your calendar, your emails, and your ERP schema.</p><p>Third: memory is the feature that actually makes an AI system feel like a system rather than a chatbot. Not just storing things — anyone can store things — but retrieving the right things at the right time, and knowing what's worth promoting to always-injected context versus what should live in archival search. Getting that hierarchy right took longer than building most of the agents themselves. It's still not perfect. But when it works, the difference is stark. Nova references a decision I made three weeks ago without me having to remind it. That's the compound effect I was chasing.</p><p>What I didn't expect was how much building Nova would change how I build everything else. The discipline of writing a tight agent spec — clear purpose, clear inputs, clear failure modes — is just good software design. I'm writing better Django views because of it. I'm making cleaner architectural decisions on AIREP because I've had to think hard about what a component is actually responsible for, versus what it's accidentally responsible for because nobody drew a boundary.</p><p>There's a broader point here that I keep coming back to: AI isn't a feature you add to software. It's a capability you develop as an engineer. The developers who are going to compound their output over the next few years aren't the ones who know how to write the best prompt — they're the ones who understand how to architect systems where AI is a reliable, testable, composable component. That's a different skill, and it takes genuine engineering investment to build.</p><p>Nova is now in what I'm calling its production orchestration phase. The agents exist; the memory system works; the routing logic is solid. The next phase is the self-improvement loop — agents that review Nova's own code, flag inconsistencies, and propose refactors. That's not a gimmick. It's a direct consequence of having a system mature enough to reason about itself. I'll write more about that when it's real, not when it's planned.</p><p>For now: 34 agents, production, no hype. Just a system that's genuinely useful and getting more so every week.</p>