Agent stack build plans · № 5 · Hermes news pipeline

Product spec

An editorial desk with one human editor — you — and a staff that works for tokens.

Who it’s for

The solo newsletter operator covering one beat, currently spending 90 minutes a day reading so they can spend 60 writing. The pipeline moves the reading and the first draft; the taste stays with you.
The devrel or comms person who has to know what moved on their ecosystem by 9am, and would rather react to three drafts than forty tabs.
The two-person editorial team where one edits mornings and one afternoons — the queue lands in a shared Slack channel and either of you can approve. (Hermes’ gateway supports group chats.)

The common thread: you already have editorial judgment and a voice. This stack amplifies both; it can’t supply either. If you don’t yet know what your bar for “worth covering” is, run the pipeline for a week and let the spike button teach you.

A day on the desk

07:00Cron run #1. Fourteen searches, eleven extracts across the watchlist. Two stories clear the bar; the desk drafts both. Queue lands in Telegram at 07:04.
07:20Over coffee: approve one draft as-is, reply “edit: cut the second section, the claim 3 hedge stays” on the other. Hermes applies the edit and returns the revision in the thread.
08:05You paste the approved pieces into your CMS yourself. (Deliberate. See below.)
12:30Cron run #2: “Nothing cleared the bar (12 searches, 8 extracts, 2 x_search).” One line. You read it in three seconds and feel fine about it.
17:30Cron run #3: one draft, fact-check flags a single-source claim. You spike it — one source isn’t enough on your beat. The spike reason goes into the thread; you fold it into the bar’s wording that weekend.
total~40 minutes of human time, all of it editing and deciding — none of it reading feeds or reconstructing sources.

What it deliberately does not do

It does not publish. No CMS credentials, no browser_type into an admin panel, ever. Drafts end at the queue and a human moves them. This is the twocents ethos and also just good sense — the failure mode of an auto-publishing news agent is public and permanent.
It does not invent quotes or colour. The house-style skill forbids unattributed quotes; the fact-checker treats any quote without a source URL as a failed claim.
It does not do deep research. The cadence is hours, the unit is a 300–600 word draft. Multi-day investigations with an archive of past reports are plan № 3.
It does not distribute. Cross-posting approved pieces to socials is a separate concern — bolt it on later if you must, but keep it out of this pipeline’s permissions.
It does not pad quiet days. The bar is explicit and the “nothing cleared it” message is a first-class output.

Cadence is a dial, and the dial costs money

“Several times a day” in this plan means three cron runs. Every run spends metered Nous Portal calls whether or not news happened. Start at two runs, add the third when you trust the bar; going hourly roughly quadruples the monitoring bill for a beat that probably doesn’t move hourly. Numbers on the traps page.