{
  "id": 3480426,
  "title": "My Cloud Run Multi-Agent Fleet Passed Its Demo. The Architecture Was Still Wrong.",
  "url": "https://urgent.news/2026/08/26/my-cloud-run-multi-agent-fleet-passed-its-demo-the-architecture-was",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-08-26T09:45:20.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/dannwaneri/my-cloud-run-multi-agent-fleet-passed-its-demo-the-architecture-was-still-wrong-1p"
  },
  "original_language": "en",
  "account": "The experiment succeeded, yet the underlying issue remained unresolved. The orchestrator detected an anomaly across three different sites within a specific timeframe, and it appropriately logged the event. Initially, the outcome appeared promising; nevertheless, I questioned the contributing factors behind its success. The answer, however, was not due to accurate logic but Cloud Run's tendency to direct both requests to the same active instance. This circumstance led the orchestrator to mistakenly believe the first instance had access to the second one's risk signature. Consequently, a correlation that should have triggered remained undetected, resulting in a flawed demonstration. My project, named VES Fleet, comprises a network of independent site-agents that conduct underground electrical surveys, analyze current flow, and calibrate their own contamination-risk thresholds based on their unique historical data. This submission aimed to compete in Google's All Things Agentic Hackathon's Fortified Enterprise Fleet track, with architectural discipline accounting for 30% of the scoring. A critical oversight was discovered when I scrutinized other early successes, such as the site list and data-provenance records. Two of the initial three sites lacked genuine underground data for calibration purposes. In response, I adjusted the project to include Bori, Choba, and Etche, which represented fully real and independently published sites. Subsequently, the bug was successfully rectified, and the state management was moved from process memory to Firestore. Each instance now reads and writes to the same shared record, eliminating the risk of misattributing successful instances to unrelated instances. To confirm the validity of the changes, I verified the system by submitting new station IDs twice during separate runs, observing a genuine correlation notice appear in both instances. In essence, while the initial demo demonstrated success, it did not guarantee the robustness of the mechanism under various conditions. The essential takeaway is that a thorough examination of the underlying system architecture is necessary to ensure reliable and accurate operation.",
  "summary": "The correlation notice fired. Three sites, same anomaly type, inside the time window. The orchestrator caught it and logged it, live, against the deployed service. Clean, first try. Then I asked myself a question I almost didn't bother asking, because the thing had just worked: why did it work? The answer wasn't \"because the logic is correct.\" It was \"because Cloud Run happened to route both…",
  "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."
}