34 Agents Is the Easy Part: Nova Hits Production Orchestration

<p>Nova — my personal AI multi-agent system — has moved into what I'm calling the production orchestration phase. There are currently 34 agents in it, spread across general chat, pipeline execution, ERP specialists, and language tooling. That number sounds impressive until you realise that spawning agents is the cheapest part of the whole exercise. Writing agent number 35 takes me an afternoon. Making agents 1 through 35 behave like a system rather than a crowd has taken considerably longer.</p><p>The honest lesson from the last stretch of work is this: an agent is a prompt, a toolset, and a scope. That's it. The interesting engineering lives entirely in the layer above — who gets invoked, with what context, in what order, and what happens when one of them is wrong.</p><p>Early on I built Nova the way most people build these things. A request comes in, something routes it to an agent, the agent does work, you get output. It works beautifully in a demo and degrades the moment you have real projects with real state behind them. The failure modes aren't dramatic. Nothing crashes. You just get subtly wrong answers — an ERP specialist reasoning about a branch it shouldn't have visibility into, a pipeline agent confidently continuing from a step that actually failed, two agents producing conflicting versions of the same artefact because neither knew the other existed.</p><p>What fixed most of that wasn't smarter models. It was treating orchestration as ordinary software with ordinary rules.</p><p>The first thing was making context deliberate rather than generous. My instinct as a developer was to give each agent everything I had — more context, better output, surely. In practice, a bloated context window makes an agent hedge. It picks up on irrelevant signals and drags them into the answer. Scoping context tightly, the same way I scope data by branch in AIREP, produced sharper results from identical prompts. Multi-tenant thinking turns out to transfer directly: an agent should only see what it has a legitimate reason to see.</p><p>The second was accepting that agents need a contract, not a conversation. When an orchestrator hands work to a sub-agent, the handoff needs a defined shape going in and a defined shape coming out. Free-form natural language between components is how you get systems that work 80% of the time and are impossible to debug. Structured in, structured out, validate at the boundary. This is not a novel idea — it's how I'd wire any service integration — but there's something about the fluency of language models that seduces you into skipping it.</p><p>The third was persistence. An agent that can't remember is a very expensive stateless function. Nova now has memory tiers — a small always-loaded profile, free-form archival notes, and searchable conversation history. The discipline there is restraint. Everything you inject into every prompt costs tokens and attention forever. Most things belong in archival storage where they can be recalled on demand, not shoved into the permanent context. I've had to delete more from Nova's core profile than I've added.</p><p>The goal I'm working toward next is the self-improvement loop: agents that review, refactor and improve Nova's own code. That's the part I'm genuinely cautious about, and I want to be straight about why. An agent that writes code for a system it also runs on needs boundaries that are enforced structurally, not by asking nicely in a prompt. Review gates, scoped write permissions, and a human in the loop for anything that touches orchestration itself. I'd rather it be slow and trustworthy than fast and surprising.</p><p>The broader point, and the reason I keep investing in this instead of just using off-the-shelf tools: AI is only a compound advantage if you own the plumbing. Anyone can call a model. The leverage comes from the system that decides when to call it, what it's allowed to know, and what it does with the result — and that system encodes your actual working knowledge. Nova isn't valuable because it has 34 agents. It's valuable because those 34 agents know how I work on AIREP, on Find a Sign, on client sites, and that knowledge accumulates.</p><p>Agents are easy. Orchestration is the product.</p>