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

<p>Nova crossed a line recently that's worth writing down. It went from being a collection of agents I'd wired together to something running in a production orchestration phase — meaning the coordination between agents is now the product, not the individual agents themselves. That distinction sounds academic until you've built both.</p><p>Right now Nova is 34 agents spread across general chat, pipeline execution, ERP specialists, and language tasks. Early on, my mental model was that if each agent was good, the system would be good. That's wrong. A pile of capable agents with no orchestration is just a pile. The hard part — the part that took the most work — is deciding which agent handles what, how context gets passed between them, and how to stop the whole thing from either stepping on itself or asking me the same question three times.</p><p>The thing that clicked is that orchestration is a routing and state problem, not an intelligence problem. The individual model calls are the easy, commoditised part. What's genuinely hard is: does this request need one agent or a pipeline of six? What context does agent four need that agent two already discovered? When an agent hits a wall, does it guess or does it stop and ask me? Getting those decisions right is the difference between a demo and something I actually rely on day to day.</p><p>A concrete example: the ERP specialist agents. AIREP is a multi-tenant Django system with branch-scoped data isolation — each client lives at the branch level. When Nova touches anything AIREP-related, it can't be casual about which tenant's data it's reasoning over. That constraint has to survive the handoff between a general chat agent that took the request and the ERP specialist that executes it. Orchestration is where that boundary gets enforced. Lose the context in the handoff and you've built something dangerous, not helpful.</p><p>The reason I'm investing this heavily in Nova isn't novelty. My whole working thesis is that AI is the primary leverage point in software engineering and business operations right now — not as a bolt-on tool, but as core competency. I'm trying to position AIREP, Find a Sign, and Nova itself so that AI is the differentiator rather than a checkbox feature. Nova is the proving ground for that. If I can't make orchestration reliable on my own system, where I control every variable, I've no business claiming it for client work.</p><p>The next milestone I care about is the self-improvement loop: agents that autonomously review, refactor, and improve Nova's own code. That's the part that feels genuinely different from ordinary automation. Most software gets better because a human sits down and improves it. The goal here is a system that closes some of that loop itself — reviewing its own agents, spotting the places where orchestration is brittle, and proposing fixes. I'm deliberately cautious about it. Autonomous code changes to the thing doing the orchestrating is exactly where you want guardrails, not enthusiasm. So the early version reviews and proposes; it doesn't merge unsupervised.</p><p>What I keep coming back to is that this is unglamorous engineering dressed up as AI. The interesting work isn't prompting. It's state management, context passing, failure handling, and knowing when the correct move is for an agent to stop and ask a human rather than confidently produce garbage. That last one is underrated. A system that knows the boundary of its own competence is worth more than one that's marginally smarter but doesn't.</p><p>Production orchestration doesn't mean Nova is finished. It means the skeleton is load-bearing now — I can add agents and pipelines onto a structure that holds, instead of rebuilding the plumbing every time. That's the milestone. The plumbing is boring, and it's the whole game.</p>