AI Systems & Agents

My AI team has seven specialists. None of them are human.

Standfirst: The move to Hermes wasn’t about a better model. It was about finally running a proper team — specialist roles, individual memory, model routing, and an escalation path that isn’t me.

Everything before this point — the copy-paste year, Ralph, Clawbot — was one agent trying to do everything. The move to Hermes was the moment the system started to feel real, because I could give each role its own identity, its own memory, and its own model.

The team

• Orchestrator — breaks work down and hands it out.
• Builder — coding and technical implementation.
• Designer — visual and user experience work.
• Researcher — finds tools, competitors and opportunities, then analyses what’s worth using.
• Copywriter — messaging and content.
• QA — testing and quality checks. Independent from the builder, always.
• Knowledge Custodian — manages all memory, tracks every document, and first-passes anything I share. It checks whether new material is better than what we have and extracts anything useful.

Model selection is handled by RouteIQ, which I built to pick the right model from the nine we currently use, per task. Not one model forced to do everything — the best tool for that specific job.

The workflow is a proper loop

Say I need an SEO analysis tool (that brief became ParSec). The researcher finds the best existing options, analyses features, identifies gaps, flags where AI could take it further. That research goes to the orchestrator, which writes the PRD. The PRD becomes a plan, the plan becomes kanban tasks, tasks get allocated to the right agents. Designer and copywriter step in where needed. QA tests everything.

Sound familiar? It’s a delivery team. The roles are the same ones I’ve managed for twenty years. That’s the point.

The escalation layer: the LLM Council

If the builder can’t resolve an issue after two proper attempts, it doesn’t hammer away and it doesn’t ask me. It escalates to the Council: five reasoning models that debate the problem and propose a solution. The Council exists because one model sometimes hits a wall another can see past.

If the Council can’t solve it, I get notified. Usually it can, and the work continues without me. (Why I make models argue before trusting them →)

When everything passes QA and UAT, it deploys to the VPS.

Where it honestly stands

I’m not at 90/10. I still make the big decisions and review final output. But I’m well past 50%, and every improvement pushes it further. The system now builds real tools — FamilyTaxi, RecipeMaster — and improves the ones I already have, rather than just talking about them.

I’m staying on Hermes because it’s the only setup I’ve found that runs this kind of specialist team with proper memory, proper task tracking, RouteIQ for model selection and the Council as an escalation layer. Stable enough that I make small weekly improvements instead of rebuilding everything each time a new model drops.

Not perfect. Working. Getting better every week. After 18 months, that’s the whole game.

Next in the series: One rambled voice note, two working apps →