AI Systems & Agents
The AI Operating System
The real product is the thing that builds the products.
Why it exists
Every app on this site was built by the same system. After a year of copy-pasting between chatbots and another six months of agentic experiments — the full story is in the writing — I stopped looking for the perfect tool and built an operating model instead. The insight that unlocked it: the harness and the model are interchangeable. The memory, the rules, the task board, the escalation paths — that’s the real architecture. Get that right and you can swap models whenever a better one appears.
The team
Seven specialist agents, each with its own identity, memory and model:
- Orchestrator — breaks work down, turns research into PRDs, hands tasks out on the kanban.
- Builder — coding and technical implementation.
- Designer — visual and UX work.
- Researcher — finds tools, competitors and opportunities; analyses what’s worth using.
- Copywriter — messaging and content.
- QA — independent testing. Never the agent that built the thing.
- Knowledge Custodian — owns all memory. First-pass reviews everything I share (articles, skills, videos) and decides whether it improves the system.
The machinery
- RouteIQ — model routing. Each task goes to the best of nine models, not one model forced to do everything. Also spreads token burn across session limits.
- The LLM Council — the escalation layer. If the builder fails twice, five reasoning models debate the problem before it ever reaches me. New ideas get scored by the council too — real demand, AI leverage, monetisation speed, competition, one-person feasibility — for 25–50p per idea. Cheaper than a wasted month.
- Ralph — the execution loop. PRD → JSON acceptance criteria → “don’t come back until it all passes.” Runs overnight; I review evidence in the morning.
- Durable memory — three tiers, with a cron that moves important context out of the hot window before it’s lost. Plus a shared vault for credentials and 30-minute heartbeats to keep agents on task.
The workflow
Brief → Researcher → Orchestrator writes the PRD → plan → kanban tasks → agents execute (Designer and Copywriter step in where needed) → QA tests → stuck work escalates to the Council → I approve → deploy to the VPS.
- BriefI define the outcome and constraints.
- ResearchThe researcher finds options, competitors, gaps and risks.
- PRDThe orchestrator turns evidence into a buildable product brief.
- KanbanThe plan becomes durable tasks with acceptance criteria.
- AgentsBuilder, designer and copywriter execute the right pieces.
- QAIndependent testing checks the work before it reaches me.
- CouncilHard failures escalate to multiple reasoning models before I intervene.
- DeployI approve the final evidence packet before production release.
Where it honestly stands
I’m not at 90/10 yet. I still own the big decisions and review final output. But I’m well past 50%, and every improvement pushes it further. This system now builds real tools and improves existing ones, rather than just talking about them.
What it proves
That the discipline is the differentiator. None of this works without proper PRDs, acceptance criteria, independent QA and escalation paths — the same things that make human delivery teams work. Twenty years of product management turned out to be the best possible training for running agents.