What Running 34 AI Agents in Production Actually Looks Like

<p>A few months ago, Nova was a side experiment. Today it's the operational backbone of how I work — 34 agents in production, handling everything from general conversation to ERP-specific queries, pipeline execution, and language tasks. I want to write honestly about what that looks like, because most writing about multi-agent AI systems is either academic or promotional. This is neither.</p><p>First, what "production" actually means here. It doesn't mean Nova is a polished SaaS product with a billing page. It means I depend on it daily. It means when an agent breaks, real work stops. It means the system has to be reliable enough that I trust it with context that matters — client data, project state, decisions I'm trying to think through. That's a different bar than a demo, and clearing it took longer than I expected.</p><p>The 34 agents are organised by function. Some handle general chat and reasoning. Others are specialists — there are ERP agents that understand AIREP's data model, branch-scoped architecture, and the kinds of questions that come up when you're building a multi-tenant Django system. There are pipeline agents that execute sequences of tasks rather than just responding to single prompts. Language task agents handle things like summarisation, drafting, and translation. Each agent has a narrow job. That's intentional.</p><p>The lesson I keep relearning is that specificity beats generality in agent design. A single "do everything" agent sounds appealing until it starts hallucinating field names in your database schema or giving you generic advice when you needed something grounded in your actual codebase. Narrow agents with tight context windows and clear mandates perform better and fail more predictably. Predictable failure is underrated — it means you know where to look when something goes wrong.</p><p>Orchestration is where the real complexity lives. Routing a user request to the right agent, deciding when to chain agents versus when a single response is sufficient, managing context so agents have what they need without being overwhelmed — none of that is solved by the model itself. You build it. It's software engineering, not prompt magic. I've written more Python glue code for Nova than I have for some client projects.</p><p>The current phase is what I'm calling production orchestration. The agents exist and work. Now I'm focused on making the system smarter about how it uses itself. The goal I'm working toward is a self-improvement loop — agents that can review Nova's own code, identify weak spots, and propose or execute refactors. That's not a gimmick. If Nova can improve its own tooling faster than I can manually iterate on it, that's a genuine compound advantage. I'm not there yet, but the architecture is pointing in that direction.</p><p>What surprised me most about building this is how much it's changed my relationship with my other projects. AIREP moves faster because I have ERP-specialist agents that hold context about the system's architecture between sessions. Find a Sign benefits from agents that help me think through marketplace design decisions without me having to re-explain the whole model every time. Sweeper Parts, which relies heavily on domain knowledge from my father's 30-plus years in the industry, is a project where I'm thinking about how to encode that expertise into agents that can assist with parts lookup and identification.</p><p>The honest summary: 34 agents in production is less glamorous than it sounds and more useful than I expected. The glamorous version is a dashboard with nodes and edges and real-time telemetry. The actual version is a system that quietly makes me faster at the work I was already doing, with occasional rough edges that I fix on weekends. It's a tool I built for myself, and it's become the clearest demonstration of something I've come to believe firmly — AI isn't a feature you add to software. It's a capability you build into how you work, at every level, until it's not separable from the work itself.</p><p>That's what I'm building toward. Nova isn't a product yet. But it's already changed what I can do, and that's the point.</p>