What Running 34 AI Agents in Production Actually Looks Like

<p>Nova is now in what I'm calling its production orchestration phase. That means 34 agents running across general chat, pipeline execution, ERP specialists, and language tasks. If you'd told me two years ago I'd be maintaining a personal AI system with that kind of surface area, I'd have assumed you were describing a well-funded startup. It's not. It's one person, running it as a backend service, wired into real work.</p><p>I want to be honest about what "production" means here, because the word gets stretched. It doesn't mean it's perfect. It doesn't mean every agent fires correctly every time. What it means is that Nova is now embedded in how I actually work — answering questions about AIREP's architecture, drafting client communication, storing and retrieving context across projects, and routing tasks to the right specialist depending on what's being asked. When it works well, it feels like leverage. When it doesn't, I fix it and move on. That's production.</p><p>The 34-agent number sounds impressive until you understand what most agents actually are: focused, single-responsibility modules that do one thing well. A blog writer agent. An ERP context agent. A memory search agent. The orchestration layer decides which ones to call and in what order. The complexity isn't in any individual agent — it's in the routing logic and the shared memory substrate that lets them pass context to each other without losing the thread.</p><p>What I've learnt building this is that the hard problems in multi-agent systems aren't the ones people talk about. It's not "can the AI reason well enough" — modern models are capable enough for most real tasks. The hard problems are: how do you keep context coherent across a long session? How do you avoid agents confidently doing the wrong thing when the prompt is ambiguous? How do you make the system observable enough that you can debug it when something goes sideways?</p><p>Memory is the part that took the longest to get right. Early versions of Nova had no real persistence — every conversation started from scratch. Useful for one-off questions, useless for anything involving ongoing projects. The current system has layered memory: a core profile that's injected into every prompt (stable facts, current goals, active projects), archival storage for free-form notes and past decisions, and episodic memory pulled semantically from conversation history. Getting the balance right — what to promote into core memory versus what to archive versus what to just let go — is genuinely an ongoing design problem.</p><p>The ERP specialist agents are where things get practically interesting. AIREP is a multi-tenant Django ERP with branch-scoped data — there's a lot of domain-specific context that a general-purpose model doesn't carry. The specialist agents are grounded in that context. They know the schema conventions, the isolation model, the business logic patterns I've settled on. When I'm working on a new AIREP module and I want to think through a data architecture decision, I'm not explaining the whole system from scratch every time. That context is already there.</p><p>One thing I keep coming back to: I don't think of Nova as a productivity tool in the way that phrase usually gets used. It's closer to a compound advantage. Each piece of context it retains, each pattern it learns about how I work, each decision it logs — that compounds. A system that knows your architecture decisions, your preferences, your active problems, and your past mistakes is qualitatively different from a general assistant that starts fresh every time. That's the bet I'm making.</p><p>There's still a lot to build. The self-improvement loop — agents that review and refactor Nova's own code — is on the roadmap but not there yet. The orchestration dashboard that shows me cross-project work-in-progress in one place is still a sketch. But the foundation is solid enough to build on. Thirty-four agents in production is a milestone. The more important milestone is that I'm actually using it, every day, for real work. That's the test that matters.</p>