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

<p>I've been building Nova — my personal multi-agent AI system — for a while now. A few weeks ago it crossed a threshold I'd call production: it's not a side experiment anymore, it's doing real work across real projects every day. Thirty-four agents across general chat, pipeline execution, ERP specialisation, language tasks, and more. That number sounds impressive. In practice, what it means is less glamorous and more useful.</p><p>Production for a personal AI system doesn't look like a launch. There's no deploy button moment. It looks like gradually trusting the system with things you used to do yourself — and noticing that it handles them better than expected, or fails in ways that teach you something. For me it started with task routing and memory. Nova knows my active projects, my values, my current goals. When I ask it something, I'm not re-explaining context from scratch. That sounds small. Over a week of work, it's enormous.</p><p>The architecture is built on a swarm model — agents with defined scopes that can hand off to each other through a central orchestrator. What I've found is that the hard part isn't building the agents. It's building the right handoff logic, and getting the memory layer right. If an agent doesn't have access to what matters, it makes generic decisions. Generic decisions are often fine. They're also often wrong in ways that are hard to spot, because they look reasonable.</p><p>This is why I've been investing heavily in Nova's memory system alongside the agent count. Every conversation, note, project update, and decision I make gets surfaced contextually. The goal isn't total recall — it's relevant recall. There's a difference. Total recall means drowning in noise. Relevant recall means the right context shows up when an agent is about to make a decision that would otherwise be uninformed.</p><p>The next phase I'm working toward is a self-improvement loop — agents that can review Nova's own code, flag patterns worth refactoring, and suggest changes. This is less about replacing my judgment and more about compressing the feedback cycle. Right now, improvements to Nova come when I notice something isn't working. A review agent changes that from reactive to proactive. I don't know yet how well it'll work in practice, but the architecture supports it and the goal is clear.</p><p>What I want to push back on is the framing that this kind of system is only for large teams or well-funded products. The leverage point is exactly the opposite. A solo developer or small team gets disproportionate benefit from AI that knows the full context of their work — because they don't have the organisational overhead to maintain that context through documentation and meetings. Nova knows I'm building AIREP as a multi-tenant Django ERP, that Find a Sign is built on a transparent no-pay-to-rank model, that I care about technical honesty over hype. That context shapes every answer it gives me.</p><p>The honest summary of where Nova is: it's useful enough that I notice when it's unavailable, and broken enough that I'm still fixing things weekly. That's probably the most accurate definition of production I know. Not perfect. Not finished. Working — and getting better.</p>