Nova Has 34 Agents. Here's What Running a Multi-Agent System in Production Actually Feels Like.

<p>Nova is in production. That sentence feels both mundane and strange to type. It's my personal AI system — a multi-agent backend that I've been building alongside everything else — and it's now past the experimental phase and into what I'm calling production orchestration. Thirty-four agents, running across general chat, pipeline execution, ERP specialist tasks, and language tooling. It's not a demo. It's the system I actually use.</p><p>I want to be honest about what that means, because the gap between "I built an AI agent" and "I run a multi-agent system" is larger than most people expect.</p><p>A single agent is straightforward enough. You give it a prompt, a set of tools, and a task. It does something. You evaluate the output. The complexity is manageable. The moment you have multiple agents that need to coordinate — hand off context, trigger each other, share state, fail gracefully — you've entered a different category of problem. It's less like writing a smart function and more like designing a distributed system, except the failure modes are weirder and the debugging is harder.</p><p>The biggest lesson so far: context is the hardest thing to get right. Each agent needs enough context to do its job, but not so much that it's drowning in noise or drifting off-task. Getting that boundary right — what each agent knows, what it can access, what it should ignore — is an ongoing calibration problem. It's not something you solve once. You tune it as the system evolves.</p><p>The second lesson: orchestration logic is real code. There's a temptation to treat the "glue" between agents as an afterthought — just route the output of one into the input of another. In practice, the orchestration layer is where most of the interesting decisions live. Which agent handles this? What happens when it fails? Do you retry, escalate, or fall back? These are engineering decisions, not prompt engineering decisions. They deserve the same rigour as any other part of the stack.</p><p>The reason I've built Nova this way — rather than just reaching for a hosted AI product — is that I see AI as the primary leverage point in how I work and what I build. Not as a tool I bolt on at the end of a project, but as a core competency that compounds over time. Every agent I add, every workflow I automate, every integration I wire up makes the whole system more capable. That's different from using a chat interface to write boilerplate. The returns are cumulative.</p><p>Right now, Nova serves as the backend AI service visible through keirantrace.com, and it feeds into how I manage work across AIREP, Find a Sign, and client projects. The ERP specialist agents are particularly interesting — they understand AIREP's data model well enough to answer questions about branch-scoped records, generate reports, and assist with data entry workflows. That's not magic; it's the result of building domain-specific context carefully and consistently.</p><p>The next phase I'm working toward is a self-improvement loop — agents that can review Nova's own code, flag issues, suggest refactors, and surface improvements. This is the part that actually excites me most, not because it's science fiction, but because it's a logical extension of what's already working. If Nova can help me manage my projects, there's no reason it can't help me improve Nova itself. The loop closes.</p><p>What I'd say to anyone building in this space: don't wait until you have a perfect architecture to ship something real. Nova has rough edges. Some agents are more reliable than others. The orchestration logic has been rewritten more than once. But running it in production — against real tasks, with real consequences when it gets something wrong — is the only way to find out what actually matters. Synthetic testing will only take you so far.</p><p>Thirty-four agents is not a goal. It's just where the count happens to sit today. The number that matters is how much useful work the system does — and that's still going up.</p>