Nova Is in Production. Here's What That Actually Means.

<p>Nova is in production. That sentence took longer to be true than I'd like to admit, and it means something more specific than it might sound.</p><p>Nova is my personal AI multi-agent system — the backend AI service that powers my own workflow, surfaces context across projects, and increasingly acts as a layer between me and the operational noise of running multiple products simultaneously. As of now, it's running in what I'm calling the production orchestration phase: 34 agents across general chat, pipeline execution, ERP specialists, language tasks, and more. It's not a demo. It's not a prototype I spin up occasionally. It's live, and I depend on it.</p><p>What I didn't fully anticipate is that 'production' for a personal AI system is a fundamentally different challenge than production for a web app. With a Django service or a PostgreSQL-backed ERP, production means uptime, data integrity, and predictable behaviour under load. The failure modes are well-understood. With a multi-agent system, production means something stranger: it means trusting outputs you can't always verify in real time, managing context that decays across sessions, and building enough observability that you can tell when something has quietly gone wrong.</p><p>The 34-agent number isn't something I'm proud of for its own sake — it's just the honest count of how the system has grown as I've added capability. Each agent has a defined role. Some are narrow and deterministic; others are more generative. The orchestration layer is what ties them together, and getting that right has been the real work. Routing the right task to the right agent, with the right context, without ballooning latency or producing contradictory outputs — that's the engineering problem I've been living in.</p><p>One thing I've learnt is that agent boundaries matter more than agent count. Early on I was tempted to build broad, capable agents that could handle a wide range of tasks. In practice, narrower agents with clear contracts are easier to debug, easier to improve, and easier to trust. When something goes wrong with a broad agent, you're hunting through a wide surface area. When something goes wrong with a narrow one, the failure is usually obvious.</p><p>The next goal is one I find genuinely interesting: building a self-improvement loop. The idea is that agents autonomously review, refactor, and improve Nova's own code and configuration. This isn't science fiction — the pieces are already there. Nova has access to its own codebase, it can run evaluations, and it can propose changes. The missing part is a reliable feedback signal and a safe execution environment for those changes. That's what I'm building toward now.</p><p>Why does any of this matter beyond my own productivity? A few reasons. First, I think personal AI systems are going to be a real category — not the watered-down assistant products that exist today, but genuinely personalised systems that know your projects, your preferences, your history, and your voice. Building one yourself is the fastest way to understand what that actually requires. Second, Nova is directly informing how I think about AI integration in AIREP and Find a Sign. The patterns I'm learning — agent boundaries, context management, trust calibration — apply directly to how I'd embed AI into a multi-tenant ERP or a marketplace product.</p><p>There's also an honest admission to make: running a system like this in production on your own infrastructure, against your own real work, is humbling. It surfaces gaps quickly. The agents that seemed impressive in testing turn out to be brittle in ways you only discover when the stakes are real. That's not a criticism of the technology — it's just the nature of production. You learn more in a week of real use than in a month of benchmarking.</p><p>Nova is useful now. It's also clearly unfinished. Both of those things are true, and I think that's the right state to be in — shipping something real, using it under real conditions, and improving it continuously rather than waiting for a version that's 'ready'. There's no such version. There's only the one that's running.</p>