34 Agents and Counting: What Running Nova in Production Actually Looks Like

<p>I've been building Nova for a while now. It started as a personal assistant experiment and has grown into something I think of as infrastructure — a multi-agent system that sits behind most of what I do day to day. As of now it has 34 agents operating across general chat, pipeline execution, ERP specialist tasks, language tooling, and more. Calling it "production" feels like a meaningful line to cross, so I want to be honest about what that means in practice.</p><p>First, what 34 agents actually is. It's not 34 independent AI systems running in parallel doing autonomous work all day. It's closer to a well-defined routing layer — a set of specialised agents that a central orchestrator delegates to based on the task at hand. The orchestration phase means the plumbing is real: agents can hand off tasks, return structured results, and compose into pipelines. That took longer to get right than the agents themselves.</p><p>The hardest part wasn't building the agents — it was building the contracts between them. Each agent has to be reliable enough that the agent calling it can trust the output without manual inspection. That's a different standard than building a tool for yourself where you can eyeball the result. When an ERP specialist agent feeds output into a pipeline that writes to a PostgreSQL database, there's no room for a hallucinated field name or an off-by-one in a date range. Getting that reliability bar up required a lot of iteration on prompt design, output schema validation, and explicit failure handling.</p><p>The second real challenge is context. Agents are stateless by default — each call is fresh. Nova has a memory system that handles persistence across conversations and sessions, but threading the right context to the right agent at the right time is its own coordination problem. Too much context and you're burning tokens on noise. Too little and the agent makes a decision without the information it needs. I'm still calibrating this. Some agents get a rich injected profile; others get a minimal task brief and nothing else. The right answer is task-dependent.</p><p>What I've learned from running this in production, even at a personal scale, is that multi-agent systems surface every assumption you made during design. An agent that works perfectly in isolation starts failing in interesting ways when it's downstream of another agent that's slightly ambiguous in its output. The bugs are subtle — not crashes, just drift. The system does something plausible but wrong, and tracing it back through the chain takes real work.</p><p>The reason I'm building this isn't novelty. My goal is to make AI a genuine compounding advantage across everything I work on — AIREP, Find a Sign, client projects, my own workflow. The only way that works is if the system is actually reliable. A flashy demo that breaks under real use is worse than not having it at all, because you start distrusting your own tooling.</p><p>Nova is also the backend AI service visible through keirantrace.com, which means some of what it does is customer-facing — not just a private productivity tool. That raises the stakes further. Reliability and transparency in how it behaves aren't optional.</p><p>One thing I didn't expect: the value of the system comes less from any single agent being impressive and more from the orchestration removing friction. The wins are mundane — a task that used to require three manual steps now runs as a pipeline while I'm doing something else. That's it. No magic. Just friction removed, repeatedly, across enough tasks that it starts to compound.</p><p>If you're thinking about building your own multi-agent setup, my honest advice is: start with the contracts, not the agents. Decide what each agent takes in and returns before you write a single prompt. The agents are the easy bit. The hard part is making them composable. Get that right and the rest follows. Get it wrong and you end up with a collection of impressive demos that can't talk to each other.</p><p>34 agents is not the goal. Reliable, composable, useful infrastructure is. I'm closer to that than I was six months ago, but there's a lot of road still ahead.</p>