Agent stack build plans · № 5 · Hermes news pipeline
What you get
A draft queue in your messages, not a firehose in your feeds. Here’s the morning delivery, as a working mock.
Three times a day, the pipeline’s cron run either says “nothing cleared the bar” — with its search and extract counts, so you know it actually looked — or delivers drafts like these. Every draft arrives with three things attached, because they’re what you’d otherwise spend your editing time reconstructing: the sources the Researcher pulled, a claim-by-claim fact-check from a separate subagent that read those sources independently, and a house-style score against the rubric in your own skill file.
Static mock of the morning queue — the three stories below are invented sample data (the first borrows a real release for flavour). Tap a collapsed draft to expand it. In the real build this arrives as messages in Telegram or Slack; a rendered page like this is what you’d share for review.
Hermes v0.19.0 lands: what “Quicksilver” actually changes
Dek: Nous Research tagged v0.19.0 on 20 July. Past the codename, the release notes reward a close read — here’s what matters if you run it as a daemon.
[draft body — 420 words, structured per /house-style: lede, two sections, further-reading]
- ① github.com/NousResearch/hermes-agent/releases — tag v2026.7.20
- ② hermes-agent.nousresearch.com/docs — changelog page
- ③–⑤ [two maintainer threads via x_search, one docs diff]
| “tagged 2026-07-20” | ✓ source ① |
| “219k GitHub stars” | ✓ source ① (checked live) |
| + 4 more claims | ✓ all matched |
Two agent frameworks are quietly merging their skill formats
Fact-checker note: claim 3 (“shipping next month”) appears only in one maintainer’s post — no second source. Draft hedges it; consider cutting.
Weekend read: what “metered tool gateways” actually cost
Style note (self-scored): two hedges in the lede, one banned word (“landscape”). Below the 85 bar — needs an edit pass before approval.
Approve / Edit / Spike are your replies in the chat thread — the pipeline’s job ends at this queue. Nothing here posts anywhere.
Why this shape
- Sources ride with the draft. The Researcher returns a numbered claims-to-URLs list; the draft cites by number. You never have to ask “where’s that from?” — the answer is attached.
- The fact-check is adversarial by construction. The Fact-Checker subagent gets the claims list and the source URLs — never the Summarizer’s prose — and re-fetches each source itself. (Hand it the draft instead and it agrees with everything; see the traps.)
- The style score is your rubric, not vibes. It comes from
the checklist in your house-style
SKILL.md— which means when the score is wrong, you fix the rubric, and the fix persists. - Quiet runs report their work. “Nothing cleared the bar (14 searches, 11 extracts)” is a feature: a pipeline that pads quiet days into content is worse than no pipeline.