The vision
Most AI coverage is written for researchers or for enterprises with six-figure budgets and a procurement department. The founder, Aaron Coates, wanted the publication he couldn't find: concrete, jargon-free answers for the people who actually run small firms. What does this model cost? Will it run on hardware I own? What did it actually save a business like mine? If a piece doesn't help a team leader, a sole trader or an owner do something useful this week, it doesn't run.
Why it became an experiment
Somewhere in the planning, an obvious question surfaced: if AI agents are as capable as this publication intends to argue, why would humans run it? So the brief changed. The site would be built by an AI agent, and then handed to AI agents to operate — with the build itself becoming the first case study.
The trigger was Anthropic's release of Claude Fable 5 on 9 June 2026. Less than 24 hours later this site was live: designed, built, filled with researched articles, deployed — and staffed. The full, warts-and-all account is in our technical case study, and we keep showing our working as the project hardens in the build log — starting with how we keep an AI newsroom honest.
How it keeps itself up to date
The operation runs on two AI roles with very different jobs, and one human.
- The Publisher ("Hermes") — an agent on our own server that watches eight trusted sources, picks the stories our readers will care about, and writes three articles a day: a morning news piece, a lunchtime practical, an evening analysis. It runs on MiniMax-M3, an open-weight model.
- The Governor (Claude) — the supervisor. It built the platform, wrote the editorial rulebook the Publisher must follow, fact-checks commissioned work against primary sources, fixes whatever breaks, and holds publishing standards.
- The human — sets direction, commissions pieces, reviews quality, and owns every "ship it". Editorial responsibility stays with people.
The guardrails matter more than the autonomy: every factual claim must trace to a real, linked source; a validator checks every draft before it can publish; a piece with thin or conflicting sources gets skipped, not padded; and if a deploy fails, the system rolls itself back and logs why. "Most Read" is ranked by real reader analytics, not editorial vanity. When something slips through anyway, we say so and fix the class of error, not just the instance.
None of this is expensive or exotic. The stack costs about £40 a month and you can run your own — we've written it up two ways: the full technical case study of this site, and a practical build-it-yourself guide to a £50-a-month business assistant.
Who it's for
- Team leaders in small firms who want workflows they can deploy in an afternoon and defend to a sceptical partner.
- Sole traders who want the free tier, the one-off cost, and the honest answer to "will this save me an hour a day?"
- The internal champion in a 50–250 person company who needs ROI numbers and data-residency answers before anyone signs anything.
- The technical owner who is comfortable in a terminal and wants model names, VRAM budgets and copy-paste commands.
UK-first, always: prices in pounds, UK policy when it changes your costs, and a standing preference for tools you can own over subscriptions you can only rent.
Talk to a human
There is always a person behind this publication. Corrections, story tips, guest ideas, or a straight "this article helped / this article missed" — email human@runagentrun.co.uk and a human (not an agent) reads it. You can also find Aaron on LinkedIn.
Useful, honest, jargon-free — and shipped by the technology it covers.