34 Agents and a System That Runs Itself: Nova in Production Orchestration

<p>A few months ago Nova was a single agent with a memory store bolted on. Today it's 34 agents, running in production, covering general conversation, pipeline execution, ERP specialists, language tasks, and more. That number isn't a vanity metric — each agent has a specific responsibility, and the orchestration layer routes work to the right one without me having to think about it. That's the point.</p><p>I want to be honest about what "production orchestration" actually means at this scale, because the phrase can sound grander than the reality. It doesn't mean Nova is autonomously running my business. It means the system is stable enough that I rely on it daily, it handles real tasks without me babysitting each step, and when something breaks I have enough observability to diagnose it quickly. That's a meaningful bar. A lot of AI projects never clear it.</p><p>The 34-agent count grew organically. I didn't sit down and architect a 34-agent system — I started with a core loop, identified the places where a single agent was doing too many unrelated things, and split responsibilities. ERP queries are fundamentally different from code generation tasks, which are different again from memory retrieval and summarisation. Keeping those concerns separate makes each agent easier to reason about and easier to improve independently. It also means failures are localised. If the ERP specialist agent misbehaves, the rest of the system keeps working.</p><p>The architecture that's emerged is roughly: a routing layer that classifies incoming requests and dispatches them, a set of specialist agents that do the actual work, and a shared memory system that lets agents pass context between each other without tight coupling. The memory layer is probably the most important piece. Without it, each agent interaction is stateless and you lose the compound value that makes a personal AI system actually useful over time.</p><p>Where it gets interesting — and where I'm spending most of my thinking right now — is the self-improvement loop. The next goal is agents that autonomously review Nova's own code, flag issues, and propose refactors. Not just assisted development where I prompt an agent to review a file, but a scheduled process that runs without me initiating it. That's a different class of system. It requires the agents to have enough context about the codebase, enough judgment about what constitutes a real problem versus noise, and enough restraint not to propose changes that break things.</p><p>I don't have that fully built yet. I'm being deliberate about it because the failure mode — an agent confidently refactoring something it doesn't fully understand — is worse than not having the feature at all. The self-improvement loop only has value if it improves things. So I'm building it incrementally: read-only analysis first, flagging and summarising potential issues, then graduated write access with human review gates before anything gets merged.</p><p>The broader point behind all of this is something I've been thinking about for a while. AI isn't useful to me as a novelty or a productivity shortcut on individual tasks. The leverage comes from compounding — systems that get better over time, that carry context forward, that reduce the cognitive overhead of running multiple projects simultaneously. Nova is my attempt to build that kind of leverage into my own workflow before packaging similar capabilities into products like AIREP and Find a Sign.</p><p>If you're building something in this space, the thing I'd push back on is the impulse to make your AI system do everything from day one. Start with one agent that does one thing reliably. Nail the memory and context layer early — it's unglamorous but it's load-bearing. And be honest with yourself about the difference between a demo and a production system. The gap is larger than it looks from the outside, and crossing it is most of the work.</p><p>More updates as the self-improvement loop takes shape. It's the part of Nova I'm most interested in — and most cautious about getting right.</p>