34 Agents in Production: What Running a Personal AI System Actually Looks Like

<p>Nova is in production. That sentence sounds cleaner than the reality, so let me be specific about what it means and what it doesn't.</p><p>Nova is my personal AI system — a multi-agent architecture I've been building alongside my other projects. It spans general chat, pipeline execution, ERP specialists, language tasks, and more. As of now it runs 34 agents. That number isn't a marketing figure; it's just the count of discrete agents that are wired up and doing real work. Some are lightweight. Some are doing heavier lifting. The point is that the system has crossed from "prototype I tinker with" into something I actually depend on day-to-day — which is a different kind of pressure.</p><p>"Production orchestration phase" is the term I'm using for where it sits right now. What that means practically: agents are being coordinated rather than just invoked individually. There's routing logic deciding which agent handles what. There's memory — notes, reminders, conversations, emails, bookmarks — being searched and surfaced across sessions. The system is starting to behave less like a collection of scripts and more like a coherent backend service. It's exposed as the AI layer behind keirantrace.com, so it's not just a local tool anymore.</p><p>What it doesn't mean: it's not self-healing, it's not fully autonomous, and it's not doing anything I'd describe as intelligent in the way the word gets thrown around. What it is doing is compounding. Each agent I add that handles a real task — drafting, searching, summarising, routing — frees up a small slice of my cognitive load. Across a full working week, those slices add up.</p><p>The next thing I'm building toward is a self-improvement loop: agents that can autonomously review, refactor, and improve Nova's own code. That's not a gimmick. The practical value is that Nova has enough context about its own architecture that it can flag redundancy, suggest consolidations, and draft improvements without me having to context-switch into "Nova maintenance mode" every time something needs tuning. Whether that actually works at the level I'm imagining is an open question. But it's the right thing to be aiming at.</p><p>One thing I've learned building this: the complexity doesn't live where you expect it. The individual agents are not the hard part. Prompt engineering a single agent to do a specific task well is tractable — you iterate, you test, you get there. The hard part is the seams. How agents hand off context to each other. How memory gets structured so that a search six weeks from now returns something useful. How you avoid the system becoming a tangle of special cases that only the original author understands. These are software engineering problems, not AI problems, and they deserve the same rigour.</p><p>I'm also conscious of a trap that's easy to fall into: building infrastructure for its own sake. Thirty-four agents sounds like a lot. Some of them probably shouldn't exist as separate agents — they could be functions, or just better prompts on an existing agent. I'm not precious about that. If a consolidation makes the system simpler and more reliable, I'll do it. The goal is a system that compounds my output, not a system that looks impressive in a diagram.</p><p>The broader point — and this is the thing I keep coming back to — is that AI is only a differentiator if it's actually integrated into how you work, not bolted on. Nova is integrated. It touches my workflow across projects: AIREP, Find a Sign, client work, this blog. It's not a chatbot I open when I'm stuck. It's infrastructure. Building it that way takes longer up front, but the compounding effect is real and it's already showing up.</p><p>More updates as the self-improvement loop takes shape. That one's going to be interesting to write about — either because it works, or because it fails in an instructive way.</p>