34 Agents and a Self-Improvement Loop: Where Nova Is Now

<p>A few months ago Nova was a personal experiment. A way to offload repetitive thinking, automate some scheduling, and explore what a personal AI system could look like if you built it yourself instead of subscribing to someone else's version. Today it's something I'd genuinely call production. Thirty-four agents, spanning general conversation, pipeline execution, ERP specialists, language tasks, and a growing set of domain-specific tools. It's running. It's useful. And now I'm thinking about what comes next.</p><p>The phase I'm in right now is orchestration. That word gets used loosely in AI circles, so I'll be specific about what it means here. It means agents aren't just executing individual tasks in isolation — they're being coordinated. One agent hands off context to another. A pipeline agent breaks a goal into steps and delegates. Results feed back into the system. The architecture has to be intentional about this, because without structure, a swarm of 34 agents is just noise with a latency bill.</p><p>Getting to this point required a lot of decisions that weren't obvious upfront. How do you scope agents so they're focused without being so narrow they're useless? How do you pass context between them without bloating every call with irrelevant history? How do you handle failure — when one agent in a chain produces garbage output, does the whole pipeline collapse, or does something upstream catch it? These aren't hypothetical questions. They're things I've had to solve, and I'm still refining the answers.</p><p>What I've landed on is a layered approach. Specialist agents stay narrow and stateless where possible. Orchestrator agents carry the broader goal and decide what to delegate. Memory — both short-term episodic and longer-term archival — is managed deliberately, not just appended to every prompt. The system knows what Nova is, what I'm working on, and what my preferences are, but individual agents don't need to know all of that. They need to know what's relevant to their task.</p><p>The next goal is more interesting and, honestly, a bit unusual: a self-improvement loop. The idea is to build agents that can autonomously review Nova's own code, identify weak spots, suggest refactors, and — under appropriate constraints — implement changes. This isn't about replacing my judgment. It's about compounding it. I write a lot of code across a lot of projects. Nova already helps me move faster. A system that can identify where Nova itself is underperforming and propose fixes closes a loop that most developer tools leave open.</p><p>I want to be honest about the risks here. Autonomous code modification is not something you point at a production system and walk away from. The loop needs guardrails — review steps, scoped permissions, clear rollback paths. The goal isn't automation for its own sake. It's building a system that gets better at helping me over time, with me still in the decision seat for anything consequential. The self-improvement loop is a productivity multiplier, not a replacement for engineering judgment.</p><p>What I find most interesting about building Nova isn't the individual capabilities — any competent engineer can wire up an LLM to do useful things. It's the architecture of a system that's genuinely personal. Nova knows my projects, my preferences, my clients, my code style. That context accumulates. The longer it runs, the more leverage it provides. That's the real differentiator, and it's not something you can buy off the shelf.</p><p>There's a broader point here about AI as a business competency. I'm building AIREP, Find a Sign, and Sweeper Parts in parallel, all as a solo operator with client work on top. Nova is how that's possible. Not because it does everything, but because it handles enough of the cognitive overhead — drafting, summarising, searching, routing — that I can stay focused on the work that actually requires me. That's the version of AI integration I'm interested in: not demos, not wrappers, but compound leverage built into how I actually work.</p><p>Thirty-four agents is a milestone, but it's not the interesting number. The interesting number is how much faster I can move next month than I can today. That's what I'm building toward.</p>