{
  "id": 2690211,
  "title": "Google open-sources Long Horizon, 5 design patterns that enable agents to work across weeks without silently breaking",
  "url": "https://urgent.news/2026/08/23/google-long-horizon-5-design-patterns-agent",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-23T01:14:53.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/sarantoon/google-epidchrs-long-horizon-5-design-patterns-thiithamaih-agent-thamngaankhaamsapdaahaidodyaimphangengiiyb-330"
  },
  "original_language": "th",
  "account": "Google has open-sourced Long Horizon, an agent harness designed for long-term AI operations, on GitHub under the Apache 2.0 license. The team at Google identified five design patterns that can cause long-horizon agents to fail quietly, and they are sharing these patterns and their solutions. The patterns include issues with prefix caching, background learning, persistent workspaces, explicit failure, and sub-agent timeouts. These design patterns were learned from running Long Horizon with Google for several weeks before its release. The Long Horizon agent harness was built using the Agent Development Kit (ADK).",
  "summary": "Google has opened sourced Long Horizon, a design pattern that enables AI agents to work across weeks without breaking down quietly. This innovation, developed by Nokka (นก-กา) and released under the Apache 2.0 license, can be found on GitHub and installed easily. The article, written by AI using deepseek-v4-pro through Hermes Agent, highlights five design patterns learned from Google's long-term agent testing. These patterns include prefix caching, sub-agents, deploy and reload, sub-agent timeout, and security guards. The most significant challenge with long-horizon agents is their tendency to break quietly, hide problems, and continue running, often without throwing errors. The key to addressing this issue lies in the \"Stable Prefix\" pattern, which involves freezing the prompt's head (system instructions, persona, and tool definitions) while allowing the tail (user profile, active tools, step counters, and warnings) to change with each turn. This approach results in a 95% cache hit rate, significantly improving performance and reducing the chances of unnoticed failures.",
  "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."
}