34 Agents Is Easy. Orchestrating Them Isn't.

<p>Nova, my personal AI system, has 34 agents now. General chat, pipeline execution, ERP specialists, language tooling — the roster keeps growing. But if you asked me what the hard part of the last few weeks has been, it wasn't building any of them. It was getting them to work together without stepping on each other.</p><p>There's a seductive story in AI right now that goes: add more agents, get more capability. It's mostly wrong. Adding an agent is cheap. A prompt, a set of tools, a scope. I can stand one up in an afternoon. What's expensive — and what Nova is now firmly in the middle of — is orchestration: deciding who owns a task, when to hand it off, what context travels with it, and how to stop three agents from all deciding they're the right one for the job.</p><p>This is the phase Nova has moved into. Not "can I make agents," but "can I make agents that route work reliably in production." The difference is enormous. In development, a messy handoff just means you re-run the thing. In production orchestration, a bad handoff means an agent confidently does the wrong task with the wrong context, and you don't find out until the output is subtly broken.</p><p>A concrete example: the ERP specialist agents that plug into AIREP. AIREP is multi-tenant with branch-scoped data — each client is isolated at the branch level, and that isolation is not optional. When an agent picks up an ERP task, the branch scope has to travel with it, unbroken, or the whole tenancy model is a lie. So orchestration here isn't a routing nicety. It's a correctness and security boundary. The agent that answers a question about one client's inventory must never be able to reach into another's. Getting the orchestration layer to carry and enforce that scope, rather than trusting each agent to remember it, has been most of the work.</p><p>The pattern I've landed on is to treat context as something the orchestrator hands down explicitly, not something agents inherit ambiently. Each agent gets exactly the scope it needs and nothing it doesn't. It's more plumbing up front, but it means an agent literally cannot act outside its lane, because it was never given the keys. That's a much better guarantee than "the prompt says not to."</p><p>The next piece I'm building toward is a self-improvement loop — agents that review, refactor, and improve Nova's own code. This is the part that makes people nervous, and honestly it should. An agent that can edit the system it runs inside is only safe if the orchestration and scoping underneath it are trustworthy. Which is exactly why I've been so stubborn about getting the boring routing and context-passing right first. You don't hand code-editing power to a swarm you can't reason about. The self-improvement loop is a reward for having done the unglamorous foundational work, not a shortcut around it.</p><p>There's a broader point here about how I think about AI in my work generally. I'm not bolting AI onto AIREP or Find a Sign as a feature to put on a landing page. The bet is that AI is the primary leverage point — the thing that compounds — in both the software and the business operations around it. But leverage only compounds if the foundation holds. A pile of agents with sloppy handoffs isn't leverage, it's a liability with good marketing.</p><p>So the milestone I'm actually proud of this week isn't the agent count. It's that Nova can now take a task, decide who owns it, pass down exactly the context that task needs, and get a clean result back — repeatably. That sounds modest written out. It's the difference between a demo and a system.</p><p>If you're building with agents, my honest advice: stop counting agents. Count clean handoffs. That number is much smaller, much harder to grow, and it's the one that actually tells you whether you've built something real.</p>