Agent stack build plans · № 3

Hermes as an infinite research intern

A daily cron job, a Docker-sandboxed scratchpad where it writes its own scrapers, and an archive of past reports so tomorrow’s briefing only contains what’s new.

Agent: Hermes by Nous Research MIT · 219k GitHub stars v0.19.0 · July 2026 Web tools: paid (Nous Portal, metered)

Hermes is Nous Research’s open-source agent — a terminal CLI plus a messaging gateway daemon, tagline “the agent that grows with you.” This plan wires three of its real capabilities into one loop: the built-in cron scheduler fires every morning; the agent researches your beat with Firecrawl-powered search and, when a source is awkward — a PDF quarterly report, a CSV dump, HTML from 2009 — writes itself a scraper as a skill and keeps it; and a three-layer memory (session search, a small pinned memory file, and a reports archive you define) means it opens each morning by checking what it already told you. The report lands in Telegram before you’re out of bed.

One thing this intern is not: a paywall-evasion machine. It exists for the awkward-but-open web — unreadable formats, not restricted ones — and the briefing you’ll paste in tells it to respect paywalls, robots.txt, rate limits, and terms of service. If a source is closed, the fix is a subscription or an API key, not a cleverer scraper.

Honest effort: an evening to the first hand-run report, then a week of mornings tuning the briefing before you trust the cron. Honest cost: every run burns LLM tokens plus metered Nous Portal calls for each web_search/web_extract — a daily schedule multiplies both by thirty. The traps page puts numbers-shaped guardrails on that.

The plan, in six parts

1What you getthe morning report, mocked up — the diff section is the whole product 2Product specwho this is for, a day in the life, and what the intern refuses to do 3Architecturecron → agent → tools → Telegram, and the three-layer memory (one layer is yours to build) 4Set it upinstall, Docker backend, gateway, cron — real commands, ~40 minutes 5The promptsstaged and paste-able: briefing, dedupe, scraper skills, schedule, human review 6The trapscost runaway on cron, the un-sandboxed local backend, and the 2,200-character “archive”

Is this for you?

Yes, if there’s a beat you already check every morning across five tabs — an industry, a research field, a set of competitors — and the sources are public but half of them are hostile to copy-paste. Yes, if you want an agent that gets better at your beat over time instead of starting cold. Skip it if your beat is adequately covered by a newsletter someone else maintains, or if a metered search bill with your name on it makes you nervous — this plan runs unattended and spends real money on a schedule. This plan deliberately does not cover multi-agent fan-out (Hermes’s delegate_task exists, but one intern is enough here), RSS monitoring (not built in — see the architecture page), or general Hermes setup as a coding agent — for terminal coding-agent plans, see № 1 in this series.