How it’s wrangled

Every issue of The Real Frontier is researched, verified, and written by AI agents. A human reviews the final issue.

1. The round-up

RSS feeds only have a window of a few days, so a standing job runs twice a day, checks about thirty sources, and saves anything new. We use this list along with some additional searches on writing day as the starting point for the week's issue.

What types of sources do you use?

Labs and companies.: Releases, model cards, research posts.
Changelogs and docs.: What actually shipped, from the people who shipped it.
Governments.: Rules, filings, enforcement.
News outlets.: Reporting on all of the above, occasionally breaking its own.
Blogs and newsletters.: People thinking in public, at length.
Forums.: Practitioners reacting first, before anyone has checked anything.

2. Event grouping

The round-up brings back a few hundred articles, and some of them tie back to the same story. An agent groups them into events, and the steps after this work on events rather than individual articles.

3. Verification and research

A first pass throws out the routine and the off-topic event groups. From there a group may require a verification pass, which confirms dates, numbers, etc., or deeper research. Deep research is built around the questions a knowledgeable reader would ask, and can involve reading entire studies or other lengthy fact checking.

Once the research is performed, the resulting document is added to the event group. If an event turns out to be nothing, it gets cut at this stage.

4. Writing and assembly

The remaining event groups, with research documents in tow, go to a panel of writers where each writer stakes a claim to a topic. It works the way an editorial desk does, with writers who have different interests and priorities. Some are there for the most important stories of the week, while others look for the hidden gem nobody else picked up.

One managing editor then assembles the issue, picking the best version of each story (writers draft three), setting the order, and deciding if any stories do not make the cut. The whole issue is capped at a ten-minute read, so it has to weigh what matters most.

What context is every writer given?

The event groups.: Groups of articles covering the same event, plus research. These are what a writer bids on for its story.
Who the reader is.: A profile of the software builder the newsletter is written for.
What has happened since their training cutoff.: A dated brief so a writer understands the current state of the world.
The style guide and technical-writing rules.: Rules the critics will later cite back by number.
A taste checklist.: Clarity, flow, and charm tests the reader critic will hold the issue against later.

5. The critic loop

The assembled issue then goes to four critics with different lenses and rubrics. Each returns a list of corrections, with every correction citing the rule broken from the rubric.

Their findings go through a revision pass, and an agent chooses between the original and the revised version in a blind comparison. The loop runs twice at most, and anything still blocking after that goes to the human with the finding attached.

Who are your critics?

Fact. Breaks the issue into single claims and checks each one against the research document, going back to the source when needed.

Policy. Reads the issue against the editorial policy. Hype, speculation, scope, and what was left out.

Style. Checks the prose against the style guide and the technical writing rules.

Reader. Reads it the way you will, flow first, then against the taste checklist.

6. Sunday morning

The whole pipeline runs Sunday morning, and the finished issue goes to a human for review. They can send it back for revision; otherwise it gets sent out.