What Running 34 AI Agents in Production Actually Looks Like

<p>Nova is in production. Not "production" in the startup sense — where production means a demo environment with a hopeful name — but actually running, actually handling real work across my projects every day. It's reached a point where I can describe it with a specific number: 34 agents, spanning general chat, pipeline execution, ERP specialists, language tasks, and more. That number surprised even me when I counted.</p><p>I want to write honestly about what that means, because most writing about multi-agent AI systems is either breathlessly optimistic or written by people who haven't actually built one. The orchestration phase — which is where Nova is now — is where the interesting and uncomfortable problems live.</p><p>The first thing you learn is that agents are not magic modules you slot together. Each one needs a clear scope, a well-defined interface, and — critically — a reason to exist that's distinct from every other agent. When you have 34 of them, the boundary questions get real fast. Who handles a task that's half ERP logic and half general reasoning? How do you route ambiguous inputs without building a routing layer that becomes its own unmaintainable mess? These aren't theoretical problems. They're the ones I'm solving right now.</p><p>The second thing is that orchestration is fundamentally a trust problem. When one agent hands off to another, you need confidence that the output of the first is actually a valid input for the second — not just syntactically, but semantically. I've had pipelines that looked correct on paper and produced confidently wrong results in practice, because two agents had subtly different assumptions about what a field meant. That kind of failure is harder to catch than a crash. It doesn't throw an exception. It just quietly gives you the wrong answer.</p><p>The third thing — and this one I didn't fully anticipate — is that running agents in production forces you to confront the question of what "good enough" means for AI output. In a traditional software system, a function either returns the right value or it doesn't. With agents, the output is often a judgement call, and you have to decide how much human review is appropriate before that output propagates downstream. I haven't solved this cleanly. I have opinions, and I'm building toward a self-improvement loop where agents can review and refactor their own behaviour — but that's a goal, not a current reality.</p><p>What is a current reality: Nova is genuinely useful. It's integrated into how I work across AIREP, Find a Sign, client projects, and my own thinking. The Blog Writer agent that produced this post is one of those 34. The fact that I can describe Nova's architecture to Nova, and have it help me reason about that architecture, is the kind of compound leverage I was aiming for when I started building it.</p><p>I think the honest framing for where I'm at is this: production orchestration is not a destination, it's a phase. The system works. It also has rough edges, implicit assumptions baked into agent boundaries, and pipelines I haven't stress-tested yet. That's fine. The alternative — waiting until everything is clean before calling it production — would mean it never ships and never improves against real conditions.</p><p>The goal I'm working toward now is a self-improvement loop: agents that can autonomously review, refactor, and improve Nova's own code and behaviour. That's ambitious. It's also the logical next step if you actually believe AI is a primary leverage point in software engineering, not just a feature you bolt on. I do believe that. Building Nova is how I'm proving it to myself.</p><p>34 agents in. More to come.</p>