What Running 34 AI Agents in Production Actually Looks Like

<p>There's a version of this post that starts with 'I'm excited to share that my AI system is now in production.' That's not this post.</p><p>Nova — the personal AI system I've been building — is now in what I'm calling a production orchestration phase. That means 34 agents are live and doing real work: general chat, pipeline execution, ERP specialists for AIREP, language tasks, and more. It's not a demo. It's not a prototype. It's the backend AI layer that runs behind keirantrace.com and increasingly handles operational work across my projects.</p><p>But I want to be precise about what 'production' means here, because the word gets stretched until it means nothing.</p><p>Production, for me, means the system is handling real inputs with real consequences. It means I'm not the safety net for every output — the agents are. It means the architecture has to hold under the kind of irregular, messy requests that come from actual use rather than contrived test cases. By that definition, yes, Nova is in production. By the definition some people use — infinitely scalable, zero downtime, enterprise-hardened — it's not, and I'm not pretending otherwise.</p><p>What's interesting at this stage isn't the agent count. 34 is just a number. What's interesting is the orchestration layer — how agents hand off to each other, how context is preserved across a multi-step pipeline, and where things break when the task doesn't fit neatly into one agent's scope. Those are the real engineering problems, and they're genuinely hard.</p><p>One thing I've learned building this: the failure mode for multi-agent systems isn't usually a single agent giving a wrong answer. It's the orchestration layer making a bad routing decision and then every subsequent agent confidently executing on a flawed premise. The error compounds quietly. That's a different class of bug than anything I dealt with in traditional software, and it requires a different approach to debugging — you're often tracing backwards through a chain of reasonable-looking decisions to find where the logic first went sideways.</p><p>The next goal I'm working toward is a Nova self-improvement loop — agents that can autonomously review, refactor, and improve Nova's own code. That's not science fiction at this point, but it's also not trivial. The scaffolding has to be right. You need agents that can read code with genuine comprehension, propose changes that are actually improvements rather than lateral moves, and operate within constraints that stop them from breaking things they don't fully understand. I don't have that yet. I have the foundation for it.</p><p>The broader point — the one I keep coming back to — is that AI isn't a feature I'm adding to my projects. It's the primary leverage point. AIREP gets smarter query handling and document processing. Find a Sign gets better supplier matching without compromising the no-pay-to-rank model I care about. Sweeper Parts gets domain knowledge encoded in a way that Wayne's 30-plus years of expertise can actually be surfaced at query time rather than locked in someone's head.</p><p>None of that is possible if I treat AI as a bolt-on. It has to be architectural — baked into how data flows, how interfaces work, how decisions get made. Nova is the system that lets me build that way across all my projects without rebuilding the AI layer from scratch each time.</p><p>I'm not going to tell you this is all figured out. The orchestration phase is called that because it's the phase where I'm figuring out how 34 agents actually work together under load, with real tasks, in ways I didn't fully anticipate when I designed them individually. Some of it is working better than expected. Some of it is revealing assumptions I shouldn't have made.</p><p>That's what production looks like. Not a launch. A reckoning with reality — and then building from there.</p>