Nova at 34 Agents: What Production Orchestration Actually Looks Like

<p>I've been building Nova — my personal AI multi-agent system — for a while now, and it recently crossed a threshold I'd describe as "production orchestration." That phrase sounds impressive, but I want to be honest about what it means in practice rather than let it float as vague technobabble.</p><p>Nova currently runs 34 agents. They span general-purpose chat, pipeline execution, ERP specialists tied to AIREP, and language tasks. The number itself isn't the milestone. I could spin up 200 agents tomorrow and call it impressive. The actual milestone is that the system is now doing real coordinated work — agents handing off to other agents, pipelines executing with context that persists across steps, and the whole thing running as a service rather than a collection of disconnected scripts I trigger manually.</p><p>Getting here required some architectural decisions I'll write about separately, but the core one was this: orchestration only works if each agent has a well-defined scope and an honest answer to the question "what does this agent not do?" The failure mode I kept running into early on was agents that were too general — they'd attempt anything, succeed at some things, and silently underperform at others. Tightening scope was the single most effective thing I did. An ERP specialist agent that only handles AIREP-related queries is more useful than a generalist that handles AIREP queries among fifty other things.</p><p>The pipeline execution layer is what turned the system from a novelty into something I actually rely on. When Nova can receive a task, decompose it, route sub-tasks to the right specialist agents, and return a coherent result — that's when it stops feeling like a demo and starts feeling like infrastructure. I'm not going to oversell it: there are still rough edges, edge cases that fall through, and tasks I wouldn't trust to the pipeline yet. But the foundation is solid enough that I'm building on top of it rather than around it.</p><p>One thing I didn't anticipate was how much the system would change the way I work on my other projects. Nova is the backend AI service visible on keirantrace.com, but its real value is internal. When I'm working on AIREP or Find a Sign, I have agents that understand those systems specifically — their schemas, their business rules, their open problems. That context doesn't live in a chat window I have to reconstruct every session. It persists. That compounding effect is what I mean when I say AI is a primary leverage point, not just a tool.</p><p>The next phase I'm working toward is a self-improvement loop — agents that can autonomously review, refactor, and improve Nova's own code. That's a longer horizon goal and it's genuinely hard to do responsibly. You don't want an agent making unsupervised changes to the system it's running inside. But the groundwork for it — reliable pipelines, scoped specialists, persistent context — is what this production orchestration phase was always building toward.</p><p>If you're building your own multi-agent system, the honest advice I'd give is: don't optimise for agent count. Optimise for scope clarity and pipeline reliability. A system with five well-scoped agents that hand off cleanly is more useful than thirty agents with fuzzy boundaries. I had to learn that by building the fuzzy version first.</p><p>Nova is also the reason I can maintain multiple active projects — AIREP, Find a Sign, Sweeper Parts, client work — without each one being a context-switch from scratch. That's the compound advantage I keep coming back to. Not AI as magic, but AI as a system that reduces the overhead of being a one-person shop across a lot of surface area.</p><p>More to come on the architecture specifics. For now, production orchestration means the thing works well enough that I'm annoyed when it doesn't — which is the truest sign something has become infrastructure.</p>