I Stopped Building an AI-first company. What is the right Hermes setup should look like

I burned out on the "autonomous AI company with zero employees" hype, dropped out for 2 months, and came back realizing my real mistake was thinking about tasks in binary — *can AI do it or not*. The unlock was a **middle tier**: tasks AI does well most of the time but that are risky enough to need a review gate. I now run Hermes as a cheap orchestrator (DeepSeek V4 Flash) that delegates all real work to Claude Code as the executor, with a Kanban board between them. Wrapping LLM calls in deterministic scripts fixed the "oops I forgot to update the status" problem. I used Openclaw for \~2 months until Anthropic restricted using their subscription for agents in April. Then I quit the race for another 2 months, mostly to get away from the noise. Every corner of the internet had someone promising a fully automatic content/money factory. Even smart, technical people fell for it. When I came back, I figured out my actual mistake. I'd been splitting every task into two buckets: **what AI can do** and **what AI can't do**. With that model you constantly find cases where the agent is unreliable and you end up babysitting it — at which point it's no better than just using Claude on your laptop. The thing I'd missed is a whole middle category. So I moved to a **3-tier system**: * **Tier 1 — Manual only.** Personal stuff I do by hand. But I still tell the agent to be proactive (e.g. "draft that email for me" even though I'll send it). * **Tier 2 — Autonomous with a review gate.** Mostly coding. Agent builds the feature, pushes to a preview branch, sends me a report + link. I review, approve, it ships. This is the tier I'd been ignoring, and it's where most of the value is. * **Tier 3 — Fully autonomous.** Scheduled/periodic jobs: weekly SEO fixes, security reviews, availability testing across my apps. Runs end to end, I just get a report. Only put things here after you've genuinely thought through the risks. **The setup that made it work:** * Host on a cheap VPS (Hetzner, 4 cores / 8GB for \~$10/mo — enough for 99% of cases). * Hermes as the **orchestrator**, running DeepSeek V4 Flash as its brain. Cheap, fast, surprisingly smart. Its job is to talk, plan, and *delegate* — never touch files/code/configs directly. * **Claude Code as the executor.** All the real work (backend, frontend, DB, security, testing, deploy) runs through the guardrails I've built up over a year. Delegating to it lets me spend my existing Max subscription limits instead of burning API money. * A built-in **Kanban board** sits between them, so every task, comment, and decision is logged and I can watch the agent work like a co-worker. **One non-obvious fix worth stealing:** when your workflows live in `.md` skill files, top models *mostly* follow them but occasionally "forget" a step ("oh you're right, I forgot that"). The fix is to wrap LLM executions in a deterministic script (Python/JS/whatever) so the important stuff — like creating the Kanban card *before* the LLM runs — is hard-coded and can't be skipped. Deterministic scaffolding around a non-deterministic core. The mental shift that fixed everything: **stop asking the model to think for you, and use the cheapest model that's smart enough to delegate.**