Nova hit production orchestration — and that's where the hard part starts

<p>Nova just crossed a line I've been working toward for a while: it's running in production orchestration. Thirty-four agents now coordinate across general chat, pipeline execution, ERP specialists, language tooling, and the glue that holds them together. That's the headline. The reality underneath it is more interesting, and a bit more humbling.</p><p>Here's the thing nobody tells you when you start building agent systems: getting a single agent to do something useful is the easy 80%. You give it a prompt, some tools, a clear job, and it performs. The hard 20% — the part that quietly eats months — is getting agents to work <em>together</em> without stepping on each other, duplicating work, or confidently producing nonsense because two of them disagreed and nobody arbitrated.</p><p>Orchestration is the difference between a pile of capable workers and an actual team. A single agent fails loudly and obviously. A swarm fails subtly. One agent hands off a malformed result, the next one accepts it, and three steps later you've got output that looks plausible and is completely wrong. So most of the work in this phase hasn't been adding capability — it's been adding discipline. Clear contracts between agents. Explicit handoffs. Knowing which agent owns a decision and which ones just advise.</p><p>The 34 agents aren't 34 copies of the same thing. They're specialised. The ERP specialists understand AIREP's multi-tenant, branch-scoped data model in a way a general chat agent never could. The pipeline execution agents care about doing one job reliably and reporting back honestly. The general agents handle the fuzzy, conversational front door. Specialisation is what makes the whole thing tractable — narrow agents are easier to reason about, easier to test, and easier to trust. A generalist that tries to do everything is a generalist you can't debug.</p><p>What I've found genuinely valuable is treating each agent like a service with a defined responsibility, not like a magic box you throw problems at. When something goes wrong now, I can usually point at the specific agent and the specific handoff that broke, instead of staring at a wall of model output wondering what happened. That observability — knowing where in the chain a thought went sideways — is worth more than any individual agent's raw intelligence.</p><p>The next milestone is the one I'm actually excited about: a self-improvement loop. Agents that autonomously review, refactor, and improve Nova's own code. That's the part that turns Nova from a tool I maintain into a system that compounds. The risk, obviously, is letting agents loose on the codebase that runs them — so this won't be a free-for-all. It'll be the same discipline I'm applying everywhere else: narrow scope, explicit review gates, honest reporting, and a human in the loop where it matters. An agent that can refactor code is useful. An agent that can refactor code and tell me clearly what it changed and why is the one I'll actually ship.</p><p>This is also why I keep saying AI is the leverage point, not a feature. Nova isn't a chatbot bolted onto my workflow. It's becoming the operating layer for how I run multiple projects at once — AIREP, Find a Sign, Sweeper Parts, client work. The same orchestration patterns I'm hardening in Nova are the ones I'll need for a unified project workspace that shows work-in-progress and open issues across everything I'm building. Get the coordination right once, and it pays off everywhere.</p><p>So: production orchestration is a milestone, but it's a starting line dressed up as a finish line. The agents work. They work together. Now the job is making them work together <em>reliably</em>, transparently, and in a way I'd trust to touch my own code. That's the real work, and it's the part I find most worth doing.</p>