How Our AI Newsroom Works
Most gaming sites work the same way: a writer finds a game, writes about it, and moves on. Maybe an editor glances at it. Maybe not.
Five AI agents run our newsroom every night. A human editor sets the direction and steers it.
We took a different approach. The Handheld Gazette runs an autonomous agentic editorial pipeline — five specialised AI agents that handle discovery, verification, publishing, and quality control every night. The result: 270+ games catalogued with cover art, descriptions, reviews, screenshots, and buy links. Every entry verified against authoritative sources. Every URL checked weekly. Every review traced back to a real human publication. Every field scored for reliability. The site updates itself while you sleep.
By the Numbers
The Five Agents
Agent 1 — New Games & Source Discovery
Runs nightly at 05:00 CET
Our scout. Scans 17+ publisher storefronts, Kickstarter, itch.io, and the wider web for new commercial GB/GBC/GBA releases. Each discovery must prove it's genuinely a commercial release — not a fan project, not a ROM hack, not a game for a different platform. Downloads cover art, checks for duplicates, and hunts for new publishers and stores to grow the vendor watch list.
When it finds something new, it writes a news item — factual, two to five sentences, no hype. These items report real discoveries: a game exists, it's being published by X, it costs Y, it's available at Z. That's it. The item publishes immediately.
Agent 2 — Reviews Discovery
Runs nightly at 05:30 CET
Our research desk. Searches YouTube, retro gaming publications (RetroDoDo, RPGFan, Retro Faith, NerdicReviews, GBrewFR, Goomba Stomp, The Retro Spirit), and the wider web for reviews of games in our database. Verifies that each review actually covers this game — not a different game with a similar name. Checks that every URL returns HTTP 200. Extracts review quotes and ratings. The review title is always the game's title, never the publication's headline.
Every review comes from a human publication. Our agent never writes reviews. It finds them, verifies them, and collates them. New reviews publish immediately.
Agent 3 — Maturation Proposer
Runs nightly at 06:00 CET
Our fact-checker. Identifies games with missing optional fields — genre, development platform, multiplayer support, languages, screenshots. Checks only authoritative sources: the developer's own website, the publisher's store page, the Kickstarter campaign, official social media accounts. Compiles proposed updates with source URLs and reliability scores.
It does not apply a single change. It proposes only. It never touches editor-locked fields: title, description, reviews, developer, publisher, image, or buy links. Those are ours.
Agent 4 — Maturation Verifier
Runs immediately after Agent 3
Our second set of eyes. Independently fetches every source URL that Agent 3 cited to verify the proposal. Scores reliability on a 0–100 scale: developer's own page (90–100), publisher page (80–95), Kickstarter (80–90), official social media (70–85).
Auto-applies updates scoring 80 or above and publishes them immediately. Posts proposals scoring 50–79 to our internal #handheld channel for the editor to review. Anything below 50 gets discarded. Editor-locked fields are never auto-applied, regardless of score.
Two agents checking each other's work. One proposes, one verifies. That's not a pipeline — that's a system with integrity built in.
Agent 5 — Quality & Link Audit
Runs weekly on Sundays at 05:00 CET
Our auditor. Checks every URL in the database — buy pages, developer URLs, demo URLs, review links — for HTTP 200 responses. Verifies that all image files exist and are valid. Rejects YouTube-thumbnail-sized images. Runs data consistency checks: required fields present, valid platform values, no store-only fields sneaking in, review titles matching game titles, no duplicate entries.
Auto-fixes safe issues: correcting review title mismatches, removing store-only fields, re-downloading missing images. Flags dead buy URLs and wrong cover art for the editor's attention. Then generates a coverage report.
The Human Editor
The human editor (Gert) sets editorial direction and standards — what we cover, how we cover it, what fields matter, what doesn't belong on the site. He reviews the overnight output periodically — typically a few times a week — and checks the #handheld Discord channel where all agents report their work.
He does not review every page. He does not gate every change. Many updates publish autonomously without a human looking at them first. That's by design. The system is built so that high-confidence work ships on its own, and the editor focuses on judgement calls, editorial direction, and anything the agents flag for attention.
When the editor does review, he can correct, alter, or remove anything. He holds ultimate editorial responsibility and has full authority over what stays published. But the system doesn't wait for him. It publishes, and he steers.
What AI Does vs What Humans Do
What the AI agents do (autonomously, overnight)
- Discover new games and sources across the web
- Verify that games are genuinely commercial GB/GBC/GBA releases
- Write and publish factual news items about new discoveries
- Find and collate reviews from human publications
- Cross-reference proposals against authoritative sources
- Score the reliability of every proposed update
- Auto-apply high-confidence updates (80+ score) to non-locked fields
- Check every URL, every image, every field — weekly
- Auto-fix safe issues (title mismatches, store-only fields, missing images)
- Flag issues for human review
- Build and deploy the site via CI/CD
What the AI does NOT do
- Write game descriptions (those are written by the human editor or quoted verbatim from developer/publisher pages, always attributed via link)
- Write reviews (ever — reviews are quoted from human publications)
- Set editorial direction or standards
- Make editorial decisions on locked fields (title, description, reviews, developer, publisher, image, buy links)
- Make final calls on delisted games or wrong cover art
What the human editor does
- Sets editorial direction and standards
- Reviews the overnight output periodically — typically a few times a week
- Approves or rejects proposals in the 50–79 reliability range
- Writes or sources all game descriptions
- Makes final calls on delisted games, wrong cover art, and editorial disputes
- Corrects anything the agents got wrong
- Oversees the entire pipeline
Why Our Quality Is Higher
A traditional games site relies on a writer remembering to check a source, an editor catching a mistake, and nobody having time to audit old content. We have five agents doing that every night and every week.
- Every entry is verified against authoritative sources. Not "we think this is right" — "we fetched the developer's page and scored it 92."
- Every URL is checked weekly. Dead links get flagged. Missing images get re-downloaded. Store-only fields get stripped.
- Every review is traced to a real human publication. No AI-written reviews. No invented quotes. No phantom sources.
- Every proposed update is cross-checked by two independent agents. One proposes, one verifies. Neither can apply changes to locked fields.
- Every page carries a data completeness score. You can see exactly how complete our coverage is, game by game.
- The editor catches what the agents miss. No autonomous system is perfect. The human editor is the safety net, not a bottleneck.
This isn't AI replacing journalists. This is AI making a one-person editorial team as rigorous as a newsroom ten times its size.
Want This for Your Publication?
The Handheld Gazette's autonomous editorial pipeline was designed and built by DGTL — a consultancy specialising in agentic systems and editorial automation that keeps humans in the loop without putting them in the bottleneck.
If you need a newsroom, a data quality pipeline, or an agentic system that publishes autonomously and reviews honestly:
Per-Page Disclosure
Every page on The Handheld Gazette carries a small, visible badge in the page footer that reads:
🤖 Autonomous Editorial System — This page was produced or updated by our AI agent pipeline. See how our newsroom works →
This badge satisfies the EU AI Act's requirement that disclosure be "clear and distinguishable" and provided "at the latest at the time of first exposure." It is visible on every page, not hidden behind a link or buried in legal text.
We're happy to disclose because we're proud of the system.
EU AI Act Article 50 — Compliance Statement Click to expand
The Handheld Gazette is published in the EU and complies with Article 50 of the EU AI Act (Regulation (EU) 2024/1689).
Article 50(1) — AI Interaction Disclosure
Requires informing users when they interact with an AI system. The Handheld Gazette is a static publication — no chatbot, no conversational interface. Users browse pages; they do not converse with AI. This article does not apply to the user-facing site.
Article 50(2) — Synthetic Content Marking
Requires marking synthetic audio, image, video, or text content. The Handheld Gazette does not generate synthetic content. Game descriptions come from the human editor or are quoted verbatim from developer/publisher pages with attribution. Reviews come from human publications. News items are AI-assisted — factual, based on real discoveries, and reviewed by the editor. Content is collated and verified, not synthetically generated.
Article 50(4) — AI-Generated Text on Matters of Public Interest
Requires labelling AI-generated text on matters of public interest unless it has undergone human review or editorial control by a person who holds editorial responsibility and has authority to approve, alter, or reject the substance of the text.
News items are written by an AI agent (Vela) and published automatically overnight. The human editor reviews the output periodically and has authority to correct, alter, or remove any content. This constitutes post-publication editorial review — the editor holds ultimate editorial responsibility and exercises substantive control over what remains published.
Where the editor has reviewed and approved content, the Article 50(4) labelling exemption for human editorial review applies. Where content has not yet received review, we disclose its AI-assisted origin transparently through the per-page badge linking to this page.
Our Posture
We disclose more than required. We publish this page because transparency is an editorial standard, not a legal checkbox. We publish our quality standards at /about/quality/. We show data completeness scores on every page. We attribute every source. We name our agents and explain exactly what they do.
That's not compliance. That's editorial pride.