34 Agents Is the Easy Part

<p>Nova — my personal AI system — is now in what I've been calling the production orchestration phase. That sounds grander than it is. What it means practically is that Nova is no longer a pile of scripts I poke at when I have a spare evening. It routes real work: general chat, pipeline execution, ERP specialist tasks, language tooling. Thirty-four agents across those areas.</p><p>Here's the thing nobody tells you about multi-agent systems: the agents are the easy part. Defining an agent is a prompt, a toolset, and a name. I could add another thirty this weekend. That's not a capability, that's a config file. The hard problem — the part that took the longest and broke the most times — is orchestration. Who answers? With what context? What do they hand off, and to whom? What happens when two agents both think they own a request?</p><p>I went through the predictable phase of trying to solve routing with a clever prompt. Give the router a list of agents and their descriptions, let the model decide. It works about as well as you'd expect in a demo and falls over the moment two agents have overlapping remits. The ERP specialist and the general chat agent both plausibly own "what's the status of that invoice logic?" The model picks one, more or less at random, and the user gets an answer that's confident and shallow.</p><p>What actually fixed it was being boring. Explicit routing rules where the domain is unambiguous, model-based routing only where it genuinely is ambiguous, and a hard rule that an agent which lacks the tools to answer must hand off rather than guess. That last one is the highest-value constraint in the whole system. A specialist that says "not mine" is worth more than a generalist that improvises.</p><p>The second thing that ate real time was memory. Every agent wanting to remember everything is how you end up with a system that's slow, expensive, and subtly wrong. Nova ended up with tiers: a small core profile that gets injected everywhere because it's genuinely stable, and a larger archival layer that's searched on demand. The discipline is in what you refuse to promote. Most facts are throwaway. A mention of a library I tried once doesn't belong in every future prompt. Getting that wrong doesn't crash anything — it just makes the system gradually dumber, which is worse, because you don't notice.</p><p>This connects to something I'm building next, which is a unified project workspace: one view across AIREP, Find a Sign, Sweeper Parts, the client sites, and Nova itself. Work in progress, open issues, recent activity. I have six or seven active work streams and the context-switching cost is the single biggest tax on my output. I don't need an AI to write my code. I need one that knows what state everything is in when I sit down, so I'm not spending twenty minutes rebuilding a mental model I had perfectly intact three days ago.</p><p>And that's the actual reason I keep investing in Nova rather than just using whatever's on offer this month. My position on AI is that it isn't a tool you bolt on — it's the primary leverage point in both engineering and running a business. But leverage only compounds if you own the thing doing the levering. Every routing rule I write, every memory tier I tune, is an asset that keeps paying. A subscription to someone else's chat window isn't. That's the bet, and it's why the next step is agents that review and refactor Nova's own code. Self-improvement loop. The system that maintains itself is worth vastly more than the system I maintain.</p><p>I'm wary of overclaiming here. There is no autonomous fleet of agents running my business. There's a well-routed set of specialists that saves me a meaningful amount of time per day and occasionally catches something I'd have missed. But it works now in a way it didn't six months ago, and the reason is unglamorous: less cleverness in the agents, more rigour in the plumbing between them.</p><p>If you're building something similar, that's the transferable lesson. Spend your time on routing, handoffs, and what you refuse to remember. The agents will be fine.</p>