{
  "id": 7895449,
  "title": "The Machine That Rejects Its Own Work",
  "url": "https://urgent.news/2026/09/16/the-machine-that-rejects-its-own-work",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-09-16T22:31:11.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/iaflux_studio/the-machine-that-rejects-its-own-work-32o9"
  },
  "original_language": "en",
  "account": "The multi-agent system, developed by iaFlux Studio, was tasked with generating content for their own website within a 35-minute timeframe. In this experiment, 69 agents were deployed, resulting in 229 minutes of agent-work compressed into just 35 minutes. At peak, 21 agents were simultaneously active. Out of 66 total deliveries, 16 texts reached the various checks, with 15 of those 16 being rejected at the first pass by the editorial check. The claims check rejected 11 texts, and the compliance check rejected 4 texts. Notably, no human intervention was required as every rejection occurred within the system itself before reaching a person. This experiment measured the opposite of AI output speed, focusing instead on how much of the generated content was discarded by the system's gates. The rejection rate of 94 percent is not considered a failure but rather the intended function of the gates to filter out unsuitable content. While the gates effectively catch what they were designed to catch, they do not replace human judgment on edge cases and introduce additional latency and maintenance costs. The rejection rate is based on one run and not a claim of zero errors, as the public documentation declares. For more detailed information on the three checks and the system's architecture, please refer to the Italian article on iaFlux Studio's website and their portfolio showcasing production systems with measured numbers.",
  "summary": "We gave a 69-agent system 35 minutes to work on the content of our own website. Sixteen texts reached the checks. The language check rejected 15 of them at the first pass, 94 percent. No human approved a text. That line reads like a story about AI writing content. It is not. It is about where the review sits. The setup The system is a multi-agent architecture running in production at iaFlux…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}