34 Agents and a Self-Improvement Loop: How I'm Thinking About AI as Infrastructure

<p>A few months ago Nova was a side experiment. Today it's a production multi-agent system with 34 agents spanning general chat, pipeline execution, ERP specialists, and language tasks. Calling it a "personal AI assistant" undersells what it actually is — and calling it production-ready infrastructure overstates where it's at. The honest answer is somewhere in the middle, which is exactly where the interesting work happens.</p><p>The framing I keep coming back to is this: AI isn't a tool you reach for. It's a compound advantage you build into the architecture of everything you're working on. A tool is something you pick up and put down. Infrastructure is something that changes what's possible across every project simultaneously. That's what I'm building toward with Nova — and increasingly, with AIREP and Find a Sign too.</p><p>What does production orchestration actually look like at this stage? Agents are no longer isolated scripts with a single purpose. They're coordinated. A pipeline execution agent can hand off to a domain-specific ERP agent, which can surface context from memory, which feeds back into the next decision. The plumbing for that is real and working. What I'm refining now is reliability — making sure the orchestration layer degrades gracefully when an agent stalls, and that the system as a whole behaves predictably under edge cases rather than just happy paths.</p><p>The goal I'm most focused on right now is the self-improvement loop: building agents that can autonomously review, refactor, and improve Nova's own code. This is the part that sounds like science fiction until you actually sit down and spec it out, at which point it becomes a fairly concrete software engineering problem. You need agents that can read a codebase, identify a specific class of improvement, write a diff, run tests, and report back — without hallucinating their way into breaking something. The challenge isn't the AI capability. It's the scaffolding around it: test coverage, deterministic evaluation, rollback mechanisms. The AI is only as useful as the harness it runs inside.</p><p>This mirrors something I've noticed across all my active projects. The bottleneck is rarely the model. It's the surrounding system — the database schema, the permission model, the way data flows between components. AIREP's multi-tenant architecture is a good example. Getting branch-scoped data isolation right in Django isn't an AI problem, it's a data architecture problem. But once that foundation is solid, AI becomes dramatically more useful inside it — because agents can reason about structured, well-scoped data rather than fighting through inconsistency.</p><p>Find a Sign is a different flavour of the same idea. The core value proposition there is customer-first discovery — transparent supplier listings, no pay-to-rank. I have a strong distrust of systems that manipulate discovery through artificial incentives. The irony is that building a genuinely fair marketplace is harder than building a pay-to-rank one, because you have to earn trust through relevance rather than just charging for visibility. AI, used well, can actually help here: better matching, smarter categorisation, surfacing the right supplier for a specific job rather than the one who paid the most. That's only possible if the underlying data model supports it.</p><p>The thread connecting all of this is that AI amplifies whatever foundation you've built — for better or worse. If your data is messy, agents will confidently produce messy outputs. If your architecture has leaky abstractions, orchestration makes them leak faster. The discipline of building clean systems matters more in an AI-augmented workflow, not less. That's probably the most counterintuitive thing I've internalised over the last year of working this way.</p><p>Nova being in production doesn't mean it's finished — it means the feedback loop is real now. Real usage, real failures, real improvements. That's the phase where you actually learn things. I'm here for it.</p>