My own agent system: it watches AI publications, respects each source’s usage rights and writes me an opportunities briefing.
21sources watched twice a dayconfig/sources.yaml
399publications on recorddata/publications.json, 26 Oct 2026
20signals, 3 themes and 4 opportunitiesknowledge/, 26 Oct 2026
65automated tests, no network or paid APIstests/
Counted in the system’s own files on 26 Oct 2026.
The problem
Hundreds of AI reports and articles come out every week. Nobody can read them all, and consulting firms’ reports can’t be copied or put through an AI.
How it’s built
Two halves that work independently: a watcher in the cloud, with no AI, and an analyst on my laptop running Claude Code. Between them, a rights policy decides what may be read.
What the machine does and what I decide
My verdicts flow back in; the AI never edits them.
One real run
The briefing of 24 October 2026, step by step.
Source: reports/2026-10-24-briefing.md and the /briefing skill.
What a briefing reads like
Summarised excerpt of the 24 Oct 2026 briefing. No consulting-firm text; opportunities and personal sections left out.
reports/2026-10-24-briefing.md
Strengthening theme
Cost per useful outcome replaces adoption as the AI metric
From 3 to 6 independent origins. The briefing’s own caveat: much of the evidence is vendor marketing.
Signals behind the theme and the week
SIG-0012
A (soon) open-weight model nears the frontier at half the price. Simon Willison
SIG-0011
Specialised small models cut inference cost by about 77%. Hugging Face blog
SIG-0013
Agent Skills becomes an open Linux Foundation standard. Anthropic
Coverage
395 publications: 30 read in full from 11 publishers and 365 by title only, 68 of them from consulting firms.
Rules it keeps by itself
Consulting firms (McKinsey, BCG, Bain, Deloitte, PwC, EY, Accenture, Oliver Wyman): title, date and link only. Not even the feed summary.
Full documents stay on my laptop; a check before every commit blocks PDFs and keys.
Undated items stay undated. Nothing counts as read unless it was read.
No API bills: the cloud half runs on free GitHub minutes and the analysis on my Claude subscription.
It’s the same method I use in the diagnosis: evidence first, then scored opportunities, and a person decides. And it shows a system like this can run on its own every day while respecting data and licences.
Tell me about a process that eats your time.
In 30 minutes we’ll see whether AI can help. No commitment.