What Running 34 AI Agents in Production Actually Looks Like

<p>Nova is in production. Not "production" in the startup sense where you slap a domain on a half-working demo and call it shipped — I mean it's running, it's orchestrating real work across multiple agents, and it's woven into how I operate day to day. That's a milestone worth being honest about, so here's what it actually looks like.</p><p>The system currently runs 34 agents. They span a few broad categories: general chat and reasoning, pipeline execution agents that chain tasks together, ERP specialists that understand AIREP's domain, and language and transformation agents for things like summarisation, drafting, and reformatting. Each agent has a defined role. None of them are general-purpose catch-alls — specificity is the point. A catch-all agent is just a worse version of a single large prompt.</p><p>The orchestration layer is what I've been building toward for a while. Individual agents are relatively easy to build. Getting them to hand off work cleanly, share context without hallucinating each other's outputs, and fail gracefully when something breaks — that's the hard part. I'm not going to pretend it's solved. It's working, but it's working the way early software works: mostly right, occasionally surprising, and requiring active maintenance.</p><p>One thing I keep coming back to is the difference between an AI tool and an AI system. A tool is something you reach for. A system is something that runs alongside you. Nova is trying to be the latter. That means it needs to understand context that persists across conversations, know which agent to route a task to, and surface information before I ask for it rather than waiting for me to prompt it. Some of that is working. Some of it is still closer to aspirational than operational.</p><p>The production phase also exposed some architectural decisions that looked fine in isolation and got weird at scale. Context management is the obvious one — when you have 34 agents, each with their own memory access patterns, you end up thinking very carefully about what gets stored, what gets injected into prompts, and what gets retrieved on demand. Too much context and you're wasting tokens and degrading quality. Too little and agents make decisions without enough information. Finding that balance is ongoing work, not a setting you configure once.</p><p>I'm also using Nova to support my other active projects — AIREP, Find a Sign, Sweeper Parts, client work. That cross-project reach is both the value proposition and the complexity driver. An ERP specialist agent that understands AIREP's branch-scoped data model is genuinely useful. But it also means I need to keep that agent's knowledge current as the system evolves, which is a maintenance burden that doesn't go away.</p><p>The goal I'm working toward — and this is in the roadmap explicitly — is a self-improvement loop. Agents that can review Nova's own code, flag issues, and propose refactors. That's not running yet. It's the next phase. I mention it because I think it's important to be clear about what's done versus what's planned. Too much writing about AI systems conflates the two, and it makes it hard to have an honest conversation about what's actually possible right now versus what's a research direction dressed up as a product feature.</p><p>What is real: 34 agents, production orchestration, daily use across real projects. What's still being built: the self-improvement loop, a unified project dashboard that surfaces cross-project work-in-progress, and tighter integration with AIREP's ERP workflows. The gap between those two lists is where most of the work lives.</p><p>If you're building something similar — a personal AI system, an internal agent framework, anything with more than a handful of specialised models — the honest advice is to get one pipeline working end-to-end before you think about scale. The orchestration problems don't show up until you have real tasks running through real agents. Abstractions that look clean in a diagram get complicated fast when an agent returns something unexpected and the next agent in the chain has to decide what to do with it.</p><p>Nova is useful now. It'll be more useful in six months. That's the honest state of play.</p>