What Running 34 AI Agents in Production Actually Looks Like

<p>Nova is in production. Not "production" as a polite way of saying I run it on my laptop and it mostly works — I mean it's the actual backend AI service powering my personal site, handling real orchestration across multiple workstreams, and actively supporting the other projects I'm building. It has 34 agents. Writing that number down still feels slightly absurd, so I want to be honest about what that actually means.</p><p>The 34 agents span a few distinct categories: general chat and reasoning, pipeline execution, ERP specialists (scoped to AIREP's domain), and language tools. They're not 34 copies of the same thing with different system prompts. Each one has a defined role in a larger orchestration layer — some agents hand off to others, some run in parallel, some only fire when a specific trigger condition is met. The architecture is closer to a workflow engine than a chatbot.</p><p>What surprised me most when moving into the production orchestration phase wasn't the technical complexity — it was the coordination overhead. When you have one agent doing a job, failure is obvious. When you have 34, failure can be silent. An agent upstream produces something slightly off, the next one in the chain accepts it without complaint, and by the time you see the output it looks plausible but is subtly wrong. That's a harder class of bug to catch than a stack trace.</p><p>So a lot of the recent work has been about observability: making sure I can see what each agent decided, why it routed somewhere, and where things went sideways. It's the same discipline you'd apply to any distributed system — you need structured logging, clear handoff contracts, and the ability to replay a pipeline from an intermediate state. The AI part doesn't change those fundamentals. It just makes the internals harder to inspect because you're dealing with probabilistic outputs rather than deterministic ones.</p><p>The next phase I'm working toward is a self-improvement loop — agents that can review Nova's own code, flag weaknesses, and propose refactors. I want Nova to be able to audit itself. That's not a gimmick. The practical value is that I'm one person running multiple projects simultaneously. If I can offload the "find the fragile parts" work to an agent that runs overnight and surfaces a prioritised list in the morning, that's real leverage. The goal isn't autonomous deployment — I'm not handing the keys over — it's autonomous review with human sign-off.</p><p>I think a lot of people in the AI space talk about "AI as a force multiplier" and then demonstrate it by generating marketing copy slightly faster. That's not what I'm after. The version I care about is: a system that knows my projects, understands the architecture decisions I've already made, tracks open issues, and can do meaningful technical work — drafting, reviewing, flagging — without me having to re-explain context every time. That's what 34 agents with persistent memory and defined roles actually buys you. Not magic. Just compound context.</p><p>Nova is also the backend for keirantrace.com, which means it's the thing prospective clients interact with when they visit my site. That creates an interesting constraint: it has to be genuinely useful, not just impressive in a demo. If it hallucinates, gives bad advice, or breaks under normal use, that reflects on me directly. That pressure is useful. It keeps the work honest.</p><p>If you're building something similar — a personal agent system, an AI layer on top of your own tooling — the thing I'd tell you is: start with the observability. Before you add the fifteenth agent, make sure you can clearly see what the first five are doing. The orchestration gets complicated fast, and "it seemed to work" is not good enough when you're building something you actually depend on.</p><p>More updates as the self-improvement loop takes shape. It's the part of this I'm most curious about — and most cautious about getting right.</p>