Nova Has 34 Agents. Here's What That Actually Means.

<p>I've been building Nova for a while now, and it recently crossed a threshold I'd call production-ready orchestration. There are 34 agents running across the system — general chat, pipeline execution, ERP specialists, language tasks, and more. If I'm honest, the number isn't the interesting part. The interesting part is what happens when you try to make a multi-agent system actually useful day-to-day, rather than impressive in a demo.</p><p>Most agent frameworks you see in the wild are demos. They chain a few LLM calls together, add some tool-calling, and call it an agentic system. Nova started that way too. The difference between that and something genuinely useful comes down to one thing: state. An agent that can't remember what it did yesterday, why it made a particular decision, or what context surrounds a request isn't orchestration — it's just a loop with extra steps.</p><p>Nova's memory architecture is where most of the real engineering effort has gone. There are distinct tiers: a core profile that gets injected into every session (stable facts, active goals, current projects), archival storage for free-form notes and lessons that don't need to live in every prompt, and episodic memory for conversation-level recall. The search across those tiers uses a hybrid of keyword and semantic matching. Getting that wrong costs you either hallucinated context or a bloated prompt that degrades generation quality — both bad outcomes.</p><p>The 34 agents break down into a few functional clusters. There are general-purpose agents for things like writing (this post is going through one), research, and task planning. There's a cluster of ERP-specialist agents that understand AIREP's domain — branch scoping, multi-tenant data, the specific workflows my clients run. There are pipeline agents that can execute multi-step tasks with branching logic. And there are agents responsible for ingesting information — emails, bookmarks, notes — and routing it to the right memory tier.</p><p>What I didn't expect is how much of the work is coordination rather than capability. Any individual agent is straightforward to build. The hard problems are: which agent should handle this request, what context does it need, how do you prevent agents from contradicting each other, and what happens when a task spans multiple agents across multiple sessions? Those aren't AI problems — they're software architecture problems. You solve them the same way you solve any distributed system problem: clear interfaces, explicit state, and logging everything so you can debug it when it breaks.</p><p>The next phase is the one I find most interesting: building a self-improvement loop. The idea is that Nova should be able to review its own code, identify patterns worth refactoring, and surface those to me — or eventually action them with appropriate guardrails. I'm not interested in fully autonomous self-modification without oversight; that's a good way to introduce subtle bugs that are hard to trace. What I want is an agent that can read the codebase, notice that three different agents are doing the same thing slightly differently, and flag it. That's a code review loop, not science fiction, and it's a tractable engineering problem.</p><p>Why build this at all? The honest answer is that AI is the primary leverage point in software engineering right now, and I'd rather understand it from the inside than treat it as a black-box service. Every pattern I work out in Nova is directly applicable to AIREP, to Find a Sign, to client work. If I can wire a self-improvement loop that actually helps me maintain a 34-agent system, that same pattern is useful for any sufficiently complex software project.</p><p>There's also a compounding effect that's hard to articulate until you've experienced it. Each agent you build teaches you something about how to build the next one better. The memory architecture I have now is the third version — the first two were too rigid, then too loose. The current one works because I stopped trying to design it perfectly upfront and started iterating based on what actually broke in production. That's not an AI insight. That's just software.</p><p>34 agents sounds like a lot. Some days it feels like a lot. But the system does things today that would have taken me hours to do manually six months ago, and that gap is only widening. That's the point.</p>