Nova Is in Production Orchestration — Here's What That Actually Means

<p>A few months ago, Nova was a single-agent chat interface I used to offload thinking. Today it's a 34-agent system in production orchestration — meaning it's not just answering questions, it's running pipelines, coordinating specialist agents, and handling real workloads across my projects. I want to write down what that transition actually looked like, because "multi-agent system" gets thrown around a lot and the reality is more specific and more interesting than the hype.</p><p>The 34 agents aren't 34 copies of the same thing. They're organised into functional clusters: general chat agents, pipeline execution agents, ERP specialists, and language task agents. The ERP specialists exist because AIREP — my Django-based ERP system — has a lot of domain-specific logic around branch-scoped data and multi-tenant isolation. A general-purpose agent doesn't know what a branch context means in that system, or why you can't just query across tenants. Specialist agents carry that context natively. That's the practical reason for the architecture, not theoretical elegance.</p><p>Pipeline execution is where orchestration actually earns its name. Instead of me manually chaining tasks — pull this data, summarise it, format it for this output, send it there — a pipeline agent handles the sequencing. I describe the goal; the system figures out the steps. It's not magic. It breaks in interesting ways when the goal is ambiguous or when an intermediate step produces unexpected output. But when it works, it genuinely compresses the time between having an idea and having a result.</p><p>The thing I keep coming back to is that I didn't build Nova as a productivity tool. I built it as infrastructure. There's a difference. A productivity tool helps you do the same work faster. Infrastructure changes what work you can do at all. When Nova is answering a client question, pulling context from a previous conversation, cross-referencing a project note, and drafting a response — that's not me working faster, that's a capability I didn't have before. That framing matters for how I think about where to invest in it.</p><p>The current goal I'm working toward is a self-improvement loop: agents that can review Nova's own code, identify weak spots, and propose or even apply refactors. That's not running yet. It's the next meaningful milestone. I'm being deliberate about it because giving an agent write access to the codebase that runs the agent is the kind of thing you want to get right before you get it fast. The architecture needs clear boundaries — what the agent can touch, what it can't, and how changes get reviewed before they go live.</p><p>What I've learnt building this is that the hard problems aren't the AI parts. Choosing a model, writing a prompt, getting a coherent response — that's the easy layer now. The hard parts are the same ones that are hard in any distributed system: state management, error handling, knowing when to retry versus when to escalate, and keeping the whole thing observable so you can debug it when something goes wrong at 11pm. Nova needs logging, it needs clear failure modes, and it needs to be something I can reason about. A black box that usually works is not useful infrastructure.</p><p>I'm also thinking about Nova's role relative to my other projects. AIREP is moving toward a productised multi-tenant SaaS ERP. Find a Sign is a marketplace built on transparent supplier discovery — no pay-to-rank, no artificial boosting. Both of those have AI integration as part of their roadmap. Nova is the system that lets me build and maintain those projects faster than I could otherwise. It's a force multiplier on everything else, which is why I treat it as a first-class project rather than a side experiment.</p><p>If you're building something similar — a personal agent system, a small multi-agent setup for your own workflow — my honest advice is to start with the plumbing. Get logging right. Define what a "task" means in your system before you start adding agents to handle tasks. The AI layer is the fun part, but it sits on top of boring infrastructure decisions that will either hold the whole thing together or let it quietly fall apart.</p><p>Nova's in production. That's the milestone. The next one is making it smarter about itself.</p>