Agent stack build plans · № 3 · part 2

Product spec

An intern with a beat, a filing cabinet, and a rule book — not a crawler with a chat interface.

Who it’s for

Anyone with a standing beat. An analyst tracking one industry, a founder watching competitors, a researcher following a subfield, an editor monitoring a policy area. The defining trait: you already do this manually, badly, most mornings.
Whose sources are public but unfriendly. Regulator PDFs, ministry CSVs, investor decks, ancient HTML, APIs with grudging rate limits. That’s the case for the scraper-skills layer — plain web_search alone doesn’t need this plan.
Who will read the report. An unread daily digest is a metered bill and nothing else. If you wouldn’t open it, don’t build it.

A day in the life

07:15The cron job in ~/.hermes/cron/ fires. Hermes loads the /research-intern briefing skill and starts the run.
07:16Sweep: web_search and x_search across the beat’s query list; web_extract on the promising hits. Each call is a metered Portal call — the briefing caps the total.
07:19The ministry stats page is a CSV behind a form. The intern runs its own krg-csv skill — a script it wrote two weeks ago, in the Docker-backed terminal — and gets a table, not a tantrum.
07:21Dedupe pass: session_search over past conversations, plus a grep of reports/. Four items match previous coverage; they go in the skipped line.
07:23Report written to reports/2026-07-23.md, then delivered to Telegram through the gateway. Housekeeping line includes the Portal-call count and a scrapers-returned-rows check.
07:40You read it over coffee and reply in the same chat: “drop the X chatter, add the EU battery-passport docket.” The intern updates its briefing skill — tomorrow is already better.

That last row is the “self-improving” part, and it’s deliberately mundane: the intern improves because its briefing, its scrapers, and its archive are all files it can edit — a skill, a scripts directory, a folder of dated markdown — with your approval gating the edits. No magic, just compounding.

What it deliberately doesn’t do

No paywall games. The intern handles formats that are awkward, not content that is closed. The briefing tells it plainly: respect paywalls, robots.txt, rate limits, and site terms; never write a scraper for a source that has said no. A paywalled source it should cover is a line item in the report (“FT reported X — headline only, no subscription”), and the durable fix is you buying access or an API key.
No RSS pretence. Hermes has no built-in RSS monitoring (as of July 2026). Feeds are covered the honest way — a fetch script the intern writes as a skill — or not at all. See architecture.
No publishing. The output is a briefing for you, not content for an audience. Auto-posting summaries of other people’s reporting is a different plan with different ethics.
No multi-agent theatre. One agent, one beat, one report. Hermes can delegate to subagents via delegate_task, but a daily digest doesn’t need an org chart.