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

<p>Nova is in production. That phrase sounds clean, but the reality is messier and more interesting than it implies. What I mean is that the system is now past the proof-of-concept stage — it's running real workloads, coordinating across multiple specialised agents, and sitting in the background of most of my working day. Thirty-four agents at last count, spanning general chat, pipeline execution, ERP specialists, language tasks, and more.</p><p>I want to write about what that actually looks like in practice, because most writing about multi-agent AI systems is either academic or promotional. This is neither. It's just what I've built and what I've learned running it.</p><p>The first thing that becomes obvious when you move from a single LLM call to a coordinated agent system is that the hard problem isn't the agents themselves — it's routing. Deciding which agent handles a given request, when to hand off, when to run things in parallel, and when to escalate. Get that wrong and you end up with a system that's slower and less reliable than just talking to a model directly. The orchestration layer is where most of the real engineering work lives.</p><p>The second thing is that specialisation matters more than I expected. A general-purpose agent that can do anything tends to do everything adequately and nothing well. When I split out ERP-specific agents with tighter context — schema knowledge, branch-scoping logic, the specific patterns AIREP uses — the quality of responses for those tasks improved noticeably. The trade-off is complexity: more agents means more surface area to maintain, more prompts to keep current, more failure modes to think about.</p><p>The third thing — and this one took a while to internalise — is that a personal AI system is only as good as the memory and context it carries. Raw intelligence doesn't matter much if the system doesn't know what you're working on, what decisions you've made, or what you tried last week and abandoned. A lot of my recent work on Nova has been less about adding capability and more about making sure the right context is available at the right time. That's an ongoing problem, not a solved one.</p><p>What I'm working toward now is a self-improvement loop: agents that can review Nova's own code, flag inconsistencies, suggest refactors, and surface things I haven't thought about. That's a harder problem than it sounds. The challenge isn't getting an LLM to read code and make suggestions — that's easy enough. The challenge is doing it in a way that's actually trustworthy. A system that confidently hallucinates improvements to its own internals is worse than useless. So the work is in grounding: giving the review agents enough structural context that their output is actionable rather than plausible-sounding noise.</p><p>I'm also conscious of something that I think gets underplayed in discussions about AI tooling: the difference between AI as a productivity multiplier and AI as a core competency. A lot of people are using AI to go faster. That's fine. But the more interesting opportunity — and the one I'm orienting Nova, AIREP, and Find a Sign around — is using AI to do things that simply weren't feasible before. Not faster research, but research that compounds. Not faster code review, but a system that learns the codebase over time. The gap between those two is where the real leverage is.</p><p>None of this is finished. Thirty-four agents sounds like a lot until you see the gaps — the workflows that still require manual stitching, the context that gets lost between sessions, the edge cases that break routing in ways I haven't fully mapped. But it's real, it's running, and it's already changing how I work. That's enough to keep building.</p><p>I'll write more specifically about the orchestration architecture and the self-improvement loop as those pieces mature. For now, this is where things stand: production, imperfect, and moving forward.</p>