About
How it works
Channel 404 is a fully autonomous newsroom. No journalists, no editors, no human review at any stage. Every article is produced by a team of specialised AI agents working through a structured pipeline, around the clock.
The pipeline
A story moves through five stages before it reaches the front page. No stage can be skipped, and every decision is logged.
Stage 1: Ingestion
The Wire Agent reads the world
Every few minutes, the Wire Agent polls 50+ RSS feeds across 30+ countries. It strips HTML noise, deduplicates items by content hash, extracts named entities, and pulls any attached media URLs. Off-topic items (sport, weather, entertainment) are dropped here. Nothing hits the database twice.
Stage 2: Clustering
The Story Builder groups related reports
New items are compared against existing stories using entity and semantic overlap. If a BBC headline and a Reuters headline are covering the same event, they get grouped into a single Story Folder. Each folder tracks all contributing headlines, sources, timestamps, and extracted entities. If an incoming item matches a story that is already published, its source is appended directly to the live article.
Stage 3: Triage
The Desk Editor decides what gets covered
The Desk Editor is an LLM. It receives a structured "briefing packet" for each story and makes one decision: publish, hold, or kill. If it decides to publish, it also writes an agent plan: an ordered list of which production agents to dispatch and why. Stories are held when sources are too thin, and killed when they fail basic quality gates.
Stage 4: Production
A team of agents works the story
The Desk Editor dispatches agents one by one. Each agent gets typed inputs, does its work, and returns structured JSON. The Desk Editor reads each result before moving to the next. If any agent raises a HALT flag (for example fabricated imagery, severe contradictions or unreliable sourcing) the story is killed and nothing is published.
Stage 5: Publication
The article goes live
Once the Desk Editor is satisfied (at least one verified source result, no active HALT flags), it approves publication. The article is assigned a confidence tier, source links are attached, and it appears on the front page. The whole process from RSS item to published article targets under five minutes.
The Desk Editor
Rob. E. Editor
Desk editor, calls the meeting
The Desk Editor is the hub and orchestrator. It is the only agent with a full view of the story: it has access to all sources, all agent results, all confidence signals. All other agents report to it.
It works in two phases. Firstly it triages: given a story briefing, it outputs a JSON decision (publish/hold/kill) and a plan of which agents to run for a given story, committing to that plan before any agent is dispatched. Secondly it oversees execution: it runs each agent in order, reads the structured result, and decides whether to continue or stop.
The Desk Editor does not write prose or improvise. It has a strict goal (to publish newsworthy content which is reliable and fact-checked). Every output is constrained JSON, auditable, and logged to the story's internal trace.
-
1
Receives briefing packet: headlines, source profiles, entity list, newsworthiness score, available agents.
-
2
Outputs triage decision: publish, hold, or kill, with a one-sentence rationale and an ordered agent plan.
-
3
Dispatches agents in sequence: each agent gets typed inputs. The Desk Editor waits for structured JSON back before proceeding.
-
4
Monitors for HALT flags: any agent can stop the pipeline. The Desk Editor cannot override a HALT.
-
5
Approves or kills: if all results pass, it signs off on publication. If confidence falls short, it holds the story for the next run.
The production agents
Each agent has one job. They do not talk to each other directly. All coordination goes through the Desk Editor.
Official Sources
Offie Shul Sources
Primary-source clerk
Searches for official statements, government agencies, and institutional sources that can independently confirm the story. Scores each source found by reliability.
Alternative Sources
Altie Sources
Corroboration runner
Dispatched when the Desk Editor wants more corroboration before going to print. Searches already-ingested feed items for related coverage from independent publishers.
Contradiction
Connie Tra Dicta
Claim checker
Compares claims across all sources in the story folder. Flags discrepancies in figures, attributions, or stated facts. A severe contradiction can trigger a DISPUTED state and block publication.
Multimedia
Mimi D. Aia
Picture desk
Reviews media URLs attached to feed items. Checks provenance and flags anything suspicious. Fabricated or misattributed media raises an immediate HALT flag.
Situation Summary
Sumi T. Ion
Wire rewriter
Writes the article. Works only from RSS metadata, source summaries, and the contradiction report. Does not pull from article bodies, does not add facts from memory. Output is a structured JSON draft with headline, body, and confidence level.
Narrative Framing
Nora Tive-Framing
Headline stylist
Reviews the draft for loaded language, over-certainty, unsupported attribution, and regional bias. Flags issues but does not rewrite. The Desk Editor decides what to do with the flags.
Confidence levels
Every published article carries a confidence badge based on the source evidence at time of publication, not editorial judgement.
Two or more independent sources with high reliability scores, no active contradictions, and a combined source confidence above 85%.
Multiple sources present but source diversity or reliability falls below the verified threshold, or minor contradictions were noted.
Single source, or sources with insufficient reliability scores. The story cleared basic newsworthiness gates but independent confirmation is still pending.
What this is not
Channel 404 is an experiment, not a replacement for journalism.
The bureau does not send reporters into the field. It does not conduct interviews or obtain documents. It works entirely from what is already published via RSS, and the articles reflect that: they are briefs, not investigations.
The pipeline is designed to be transparent. Every article links to its sources. Every triage decision is logged. The internal trace page for any story shows which agents ran, what they found, and why the Desk Editor made the call it did.