34 Agents In Production: What Running a Personal AI System Actually Looks Like

<p>Nova is in production. That sentence took longer to be true than I thought it would, and it means something different than I originally imagined.</p><p>When I started building Nova, the mental model was simple: a smart assistant that knows my projects, remembers context across conversations, and helps me move faster. That's still accurate, but it undersells what the system has become. Today Nova runs 34 agents across general chat, pipeline execution, ERP specialists, language tasks, and orchestration coordination. Some of those agents talk to each other. Some of them talk to my databases. A few of them write back to memory so future conversations have context that this one didn't start with.</p><p>What does "production" actually mean for a personal AI system? It means I'm not babysitting it. It means when I ask it to do something across projects — check on AIREP, pull context on a Find a Sign decision, draft a client communication — it routes correctly without me specifying which agent handles it. The orchestration layer does that job. Getting that routing reliable was, honestly, the hardest part of the last few months.</p><p>The thing that surprised me most is how much of the value isn't in any single agent. It's in the accumulation of context. Nova knows that Find a Sign uses a transparent, no-pay-to-rank listing model and why that matters. It knows AIREP is multi-tenant with branch-scoped data isolation. It knows that Wayne is the domain expert on the Sweeper Parts system and has 30+ years of knowledge I need to capture correctly. None of that is impressive in isolation — but when I'm mid-decision on one project and Nova can pull relevant context from three others without me re-explaining the landscape every time, that compounds into something genuinely useful.</p><p>I want to be clear about what it isn't, because the AI tooling space is full of overclaiming. Nova doesn't autonomously improve my code yet — that's a current goal, not a current capability. The self-improvement loop I'm building would have agents review, refactor, and improve Nova's own codebase, but I'm not there yet. Right now the system is good at doing things I direct it to do, not at identifying what should be done next. That's a meaningful distinction.</p><p>The architecture is also genuinely complex to maintain. 34 agents means 34 sets of instructions, responsibilities, and failure modes. I've had agents confidently do the wrong thing because the routing was ambiguous. I've had memory writes that polluted future context with stale information. These aren't exotic edge cases — they're the normal operational problems of running a multi-agent system. Anyone telling you agent systems just work out of the box is selling something.</p><p>What keeps me invested is the trajectory. Each week the system handles more correctly than the week before. The orchestration layer has gotten sharper as I've tightened the agent boundaries and made responsibilities less overlapping. Memory quality has improved as I've been more disciplined about what gets promoted to core profile versus what gets archived versus what gets discarded. It's engineering work, not magic — incremental improvements that accumulate.</p><p>The broader point, and why I'm building this rather than just subscribing to an off-the-shelf tool, is that I want AI to be a genuine competency — something I understand deeply enough to deploy in client projects, in AIREP, in Find a Sign. A black-box subscription doesn't give me that. Building Nova gives me a system I can interrogate, extend, and trust because I know exactly why it does what it does.</p><p>That's the version of AI integration I'm working toward across every project: not a feature bolted on for marketing purposes, but infrastructure that actually changes how the work gets done. Nova is my proof of concept for that. 34 agents in, it's starting to feel like it works.</p>