We argue against every story before you see it.
Most news services show you everything that loosely matches a keyword. Agentic News runs an LLM Devil's Advocate that actively argues against publishing each story — geographic bleed, keyword overlap, vendor content, bio profiles, job postings, and stale news all get rejected before anything reaches a reader.
The log below is real. It updates as the validator runs.
Two stages. Skeptical first.
Before any content is generated, an LLM judge looks at the article and the user's monitoring agent and asks: is there a legitimate reason NOT to show this?It checks for eight specific failure modes — geographic bleed, keyword overlap with different domain meaning, broad-topic matches that miss the specific angle, promotional content, sub-domain mismatches, manifesto/about-page content, job listings, and vendor blog posts that lack original reporting.
The judge writes a one-line reason for its verdict. Skeptical by default — when uncertain, reject.
Only stories that survive Stage 1 reach the second stage, where hooks, previews, and digest priorities are generated. Separating judgment from generation matters: when an LLM is asked to write a hook, it presupposes relevance. By running the verdict step first, with no generation pressure, the system is free to say no without having invested work in the story.
For accepted stories, the judge's reasoning is shown to readers as “Why I picked this” — the same explanation, kept on the story.