I mention "Hermes" often enough on this site that it deserves its own explanation, properly, rather than a name-drop. It's not a product I'm selling you. It's the system I built for myself, and it's the honest answer to "how does any of this actually work under the bonnet."
Pick a single AI provider and build your whole operation around it, and you've inherited their pricing, their outages, their roadmap, and whatever they decide to change next quarter. I wrote about the cost side of this in an earlier note — the short version is that running every task through one expensive model is the most common way small businesses quietly overspend. The deeper version is about resilience, not just cost: if your entire business depends on one company's model staying available, cheap, and good, you don't actually control your own operation.
Think of it as the office, not as an AI itself. Hermes doesn't do the work — it's the portal that routes work to whichever specialist is right for the job. I use it across several providers (Gemini, Claude, Grok, Qwen, and others as they're worth adding), the same way a real office has different staff for different jobs rather than one person doing everything badly.
The models are the specialist staff. The agents are staff who actually do a job — draft a reply, check a database entry, summarise a document. The shared memory across the system is the point that makes it an office and not just a pile of separate tools: what one agent learns or produces, another can pick up, instead of starting from nothing every time.
Concretely, not abstractly: a piece of work comes in — say, a monthly business review needs drafting. A cheaper, fast model handles the grunt work first (pulling together what happened, a rough first draft). A stronger model is brought in only for the part that actually needs judgement — which angle is honest, which is overclaiming, what to cut. The draft then sits, held, until a human — me — reads it and decides whether it's fit to publish.
That last step isn't a formality. Nothing Hermes touches posts to social on its own, ever — a human posts, because the ban risk of an agent posting badly is real, and because "agents draft, humans decide" isn't a slogan here, it's the actual mechanism. Publishing to my own site happens only with the same explicit approval.
Partly cost control, partly not wanting to be structurally dependent on one vendor's tooling decisions, and partly — honestly — because building it is how I actually keep up with a field that's moving weekly. I'd rather have hands genuinely on the plumbing than a polished dashboard I don't understand the inside of. That's also, not coincidentally, the same standard I hold anything I'd build for your business to.
I want to be straight about scope. Hermes today reliably does routing, drafting, and holding work for review — that's real, working, and how the content on this site actually gets made. A fully autonomous version — agents independently researching and running an entire content strategy end to end — is a direction I'm building toward, not something live today. If anyone tells you their AI system is fully autonomous and unsupervised, that's exactly the kind of claim worth being sceptical of.
This is the same discipline behind anything I'd build for you — the right model for the job, cost-aware, and nothing goes live without a human checking it first.
See how it works