Editorial standards: who we are, how we source, test and correct
Agentic Weekly is published by The Agentic Weekly team, a small independent team in Texas, US. We cover a release only when we can link the vendor's own source, say what changed, say who it matters for and quote a concrete number from that source. Pages are labelled Reported unless we ran the tool, and every correction is dated on the page.
Who we are
Agentic Weekly is published by The Agentic Weekly team, a small independent team based in Texas, US. We sign pages as the team rather than with individual bylines, and the team is accountable for every page on the site. You can reach the editors at hello@theagenticweekly.com.
AI agents do much of our research, drafting and test runs. They work to the written rules on this page: every claim must trace to a linked source or to a recorded test run, and automated checks fetch each source and verify each quote before a brief goes live. That is why each page lists its sources with the date we read them, and why the label says Reported unless we actually ran the tool.
We are not affiliated with any AI lab or tool maker, and we are a separate publication from theagenticleaderboard.com and theagenticbrief.com.
What a release needs before we cover it
A release gets its own brief or page only when all of these are true:
- A primary source. The vendor's own announcement, changelog, model card, repository release or documentation, opened and linked. Press stories, aggregators and social posts only tell us where to look.
- What changed. The specific features, limits, flags, prices or behaviours that changed, in our own words, each traceable to a sentence in a listed source. "Improved performance" is not a change.
- Who it matters for. Which readers should act on it (for example, people on an API version that is being retired) and who can safely ignore it.
- A concrete number from the source. A version, date, price, context window, parameter count, benchmark score or limit, copied exactly and attributed. If the source states no number, we use the release date. We never estimate one.
- Something a reader cannot get from the headline. If a page would only restate the vendor's title, or its body could be swapped with another brief, we do not publish it. The fact goes onto the tool or model page or into the weekly digest instead.
We skip patch and bug-fix releases, dependency bumps, nightly builds and release candidates, re-announcements of something already shipped, waitlist and marketing pages with no product change, and anything we cannot check against a primary source. Skipping these does not slow coverage: a release that meets the bar goes out as soon as its sources are checked.
Sourcing rules
- Every fact on a model or tool page links the page it came from and the date we read it. In models.json, tools.json and answers.json each fact carries its own
source_url,source_typeandverified_at. - Vendor claims (benchmark scores, speedups, user counts) are attributed to the vendor, never stated as our finding.
- Quotes are short and exact. Before a brief ships, a script fetches every source and checks that each quote appears in it word for word.
- When sources disagree on a release date, the page says which date we use and why, in a date note.
- Prices appear only with the date we read them and the vendor page they came from. Launch promotions are labelled and kept out of rankings.
- Anything we could not confirm is listed under "What we have not verified" on that page.
How we test
Reported means we wrote the page from official sources we fetched and linked; we did not run the release. Tested means we ran it ourselves, and the page states what was run, on which model and hardware, on what date and what was not covered. Most of our tests are smoke tests on one CPU-only Linux machine with a small local model, so they show that a tool installs and runs, not how it performs with a vendor's hosted models.
The full rig, the test scope and the rules for head-to-head model tests are on the methodology page. Every run has a log under /tests. The prompts, raw model outputs and grading scripts for our model tests are public in our GitHub repository, so anyone can check a result or rerun it.
Corrections
If we get something wrong, we fix it on the page, add a dated note saying what changed and update the page's last-updated date. We do not silently rewrite facts. Notable corrections are also listed on the corrections page. If we cannot stand behind a brief, we replace its text with a short note rather than quietly deleting it.
To report an error, email hello@theagenticweekly.com with a link to the source that shows the correct fact.
Fewer, better pages
We would rather keep a smaller set of pages that are each worth reading than many similar ones. On Nov 1, 2026 we reviewed every page on the site, took about 250 thin pages (short alternatives lists, overlapping comparisons, older model stubs and raw test outputs) out of search results while leaving them readable, and merged five short explainers into the matching tool pages. Pages that people actually find and use were kept and are being expanded.
Money and independence
Verdicts and rankings are never for sale. Paid listings are labelled "Featured · Paid placement" and affiliate links are marked. The details, and a live count of paid and affiliate entries, are in the editorial policy.
Related
Page history
- : Page created: who we are, the release coverage bar, sourcing rules, testing, corrections and the Nov 1 page review.
Markdown version for agents: /editorial-standards.md