An autonomous newsroom. Four chained morning routines find what people search for, research it against ranked sources, write a cited article, run it past three independent reviewers, and publish. Four businesses, one system.
What I did
Pipeline architecture
Editorial standards
14 agent roles
Ops across 4 businesses
See published output ↗
Every stat traceable to its original source
70
Articles live across two blogs alone
952
Source citations placed
66%
Of data points from tier 1 primary sources
14
Specialized agent roles
The problem
Anyone can generate an article. The hard part is an article that survives scrutiny from readers, from Google, and from the AI engines that quote sources.
The goal: an editorial standard the pipeline itself enforces.
What I built
- The whole newsroom, mapped in the chart above: 14 specialized agent roles from keyword discovery through publish, chained across four scheduled morning routines.
- The editorial standard as code, a 4 tier source taxonomy, a 60% tier 1 gate, and style limits a tester can fail.
- The publishing rails, which write straight to four live company blogs. 70 articles shipped this way.
- The human gate, the one boundary the system may never cross on its own: contacting a person.
Decisions that mattered
- 14 roles instead of one monolith. Agents exchange file paths, never raw content, so I can inspect every artifact.
- The writer fixes its own bug. A failed review returns to the agent that wrote the piece, with context. Three rounds, then it escalates to me.
- Written for the machines that quote it. Charts are accessible SVG with hidden data tables, and structured data marks what a search engine may quote.
- Taste became testable rules. Stat density, table length, and banned styles are hard limits a tester can fail.
A source that fails verification gets downgraded and qualified, never dropped without a trace.