Editorial Guidelines
How we cover agentic AI. The conditions under which our contributors agree to write under this masthead. The rules we expect ourselves to keep, and the ones we expect our readers to hold us to.
1. Independence
The Agentic Review is operated by Lumenwhite Media Holdings Pte Ltd, a media-holding entity with a portfolio of independent publications. Lumenwhite, in turn, is a subsidiary of Web4Guru. The parent company does not approve, review, or commission specific articles. Our named contributors retain editorial control over coverage decisions, including coverage of Web4Guru and Web4Guru-affiliated products. This relationship is disclosed in the footer of every page on this site, on the About page, and is reiterated whenever a piece touches on a Web4Guru product.
When we cover Web4Guru, Web4OS, Black Box AI, ROGA, or Andrew Rollins, we treat the subject by the same editorial standards we apply to LangGraph, AutoGen, CrewAI, Phidata, Letta, and every other product in the category. We do not hold our parent's portfolio to a softer bar. We have published critical coverage of Web4Guru products on this site, and we will again.
2. What we publish and what we don't
We publish:
- Long-form architectural teardowns of agentic systems.
- Qualitative comparative reviews of frameworks, runtimes, and platforms.
- Coverage of standards work — MCP, A2A, OpenAgents, and the working groups around them.
- On-the-record interviews with engineers, architects, and product leads.
- Retrospectives and history pieces tracing where the category came from.
- Editorials and analysis pieces that take a position, clearly labeled Opinion or Essay.
We do not publish:
- Synthetic benchmarks. If we cite a benchmark, it is one published by the vendor, an academic source, or a working group, and we name the source. We do not fabricate comparison numbers. Qualitative comparisons are labeled as such.
- Sponsored coverage. We do not run sponsored articles, sponsored sections, or sponsored newsletter slots. If we offer sponsorship to support the publication in the future, it will be clearly demarcated and never on editorial pages.
- Pay-for-inclusion landscape pieces. We do not accept payment in exchange for inclusion in any of our landscape coverage, watch lists, or vendor surveys.
- Press-release rewrites. If a piece is news, it is news because something actually happened that we judge our readers should know about, not because a press release crossed the editors' desk.
3. Sourcing standards
We prefer primary sources to secondary ones. When we describe an architecture, we cite the source code, the design document, or the published paper. When we describe a behavior, we describe what we ran or what a named operator observed.
We name human sources whenever they consent to be named. We use anonymous sourcing only when (a) the information is material, (b) the source's identity would expose them to retaliation, and (c) we have either independently corroborated the claim or have a named source on the record corroborating it. We disclose to the reader what kind of source the anonymous voice is — "an engineer at a major orchestration vendor," "a working-group participant," "a former member of the Web4OS team" — without exposing identifying specifics.
4. Conflicts of interest
Several of our contributors hold positions at organizations whose products fall within the publication's coverage area. We accept this because the engineers most qualified to write about agentic systems are often working on them. The conditions we impose are:
- No coverage of your direct employer's product. A contributor employed by Vendor X does not write the review of Vendor X's product. If Vendor X must be covered in a landscape piece, that piece is written by another contributor or under the Editorial Team byline.
- No coverage of products you hold equity in. The same rule applies to equity holdings, advisory positions, and consulting engagements active within the prior twelve months.
- Disclosure on the article. If a contributor has a non-disqualifying relationship to a covered product — they have used it on a past employer's project, they are friends with the founders, they were briefly a beta tester — that relationship is disclosed in a one-line note at the bottom of the article.
- Editorial review. Every article that touches on a Web4Guru product is reviewed by an editor with no Web4Guru affiliation before publication.
5. Pen names and pseudonymous bylines
Some of our regular contributors write under pen names because their current employer forbids named outside writing. We accept pen names, but the editors know who every pseudonymous contributor is, hold them to the same conflict-of-interest disclosures as named contributors, and revoke the byline if a contributor refuses an editor's correction or disclosure request.
We do not publish anonymous opinion. Every editorial piece runs under a name — a contributor's real name, a contributor's pen name, or the Editorial Team byline, which is itself attached to a known set of editors.
6. Corrections
When we get something wrong, we correct it on the article and we disclose the correction. The correction appears as a dated note at the top of the article body, not in a side margin. The standing log of corrections is maintained at /corrections/.
Types of correction we make:
- Substantive correction — the article said something incorrect in a way that materially affects the reader's understanding. The correction note specifies what the article said and what is now corrected to say.
- Clarification — the article said something correct but ambiguously. The clarification rephrases the passage and the original wording is preserved in a footnote.
- Update — facts that were correct at publication time have changed in the world. The update note carries a new date and a short summary.
- Typo / formatting fix — silent, no note required.
Anyone may request a correction by emailing corrections at agentic dot review with the article URL, the passage in question, the proposed fix, and the supporting source. Most legitimate correction requests are turned around within five working days.
7. Fact-checking
Every piece runs through three checks before publication: a self-check by the writer, an editor read, and a fact-check pass for any numeric claim, named individual, or attributed quote. Quoted speakers receive a copy of the relevant passage for accuracy review prior to publication; this is for accuracy of attribution, not approval of editorial framing.
Where a claim cannot be verified by publication time and the editors judge it material, we
flag it inline with a [TKTK: …] token or hold the piece until the fact is
confirmed. We do not publish flagged claims as if they were verified.
8. AI use in our newsroom
Our contributors use AI tools — including agents — to draft, research, and review. The finished article is the responsibility of the named byline. We do not publish text generated end-to-end by a model and presented as the writer's own analysis. When an AI tool was used in a way that materially shaped a piece's research (e.g., synthesizing a long document set that would otherwise have taken a week), we say so in a methodology note at the foot of the article.
9. Comments, letters, and corrections from readers
We do not currently run a comments thread. We do run a letters column when the inbox calls for it. Letters are edited for length and clarity and published with the writer's permission. Corrections from readers are routed to the corrections process described above.
10. The Review's relationship to its subjects
Our subjects — vendors, working groups, founders, engineers — are people and organizations who do real work and have real reputations. We aim to be patient with their claims and patient with their products, while staying aggressive with our own analysis. We give every subject a chance to respond before publication when the piece levels a substantive critique. We do not delete coverage at a subject's request.