Agent stack build plans · № 5
Hermes as a live news-to-content pipeline
Cron jobs watch your beat, three delegated roles turn stories into drafts, and the queue lands in your Telegram — with a human between the pipeline and the publish button.
Hermes
is Nous Research’s open-source agent — a terminal CLI plus a
hermes gateway daemon that lives in your messaging apps. This plan
wires it into a working editorial desk for one beat: scheduled cron
runs search the web and X for movement on your topics, and when a story
clears your bar, a Researcher, a Summarizer, and a Fact-Checker
— spawned as isolated subagents via delegate_task, each on a model
suited to its job — turn it into a draft in your house voice. Drafts arrive in
Telegram or Slack several times a day. You approve, edit, or spike. Nothing
publishes itself.
What you end up with: a morning queue of two or three drafts, each carrying
its source links, a claim-by-claim fact-check table, and a house-style score —
plus honest “nothing cleared the bar” messages on quiet runs. The
voice lives in two real Hermes mechanisms: a short SOUL.md persona
and a house-style skill (SKILL.md) you invoke at drafting time —
so the voice survives restarts, model swaps, and your own drift.
Honest effort: an evening to first draft, then a week
of supervised runs before you trust the cadence. Honest cost: the web
tools (web_search, web_extract, x_search,
the cloud browser) run through the paid, metered Nous Portal tool
gateway — and a pipeline’s whole job is calling them on a schedule.
The traps page does the multiplication before your first
invoice does.
The plan, in six parts
Is this for you?
Yes, if you already spend an hour-plus a day reading feeds to write a newsletter, a beat blog, or devrel updates, and the bottleneck is drafting, not judgment. Skip it if you wanted a machine that publishes while you sleep — this plan never connects the pipeline to a CMS, on purpose; the human approval step is the point, not a limitation. Scope cuts: this is a news cadence (hours), not deep multi-day research — that’s № 3, the research intern; and it drafts, it doesn’t distribute — pushing approved pieces to your channels is a different job. One correction up front, because other guides get it wrong: Hermes has no built-in RSS reader (as of v0.19.0). Monitoring here is cron + search + extract, honestly labelled, with an optional RSS-to-webhook bridge you build yourself.