{
  "id": 12708118,
  "title": "Part 1: Make your AI agents boring: the determinism layer",
  "url": "https://urgent.news/2026/10/07/part-1-make-your-ai-agents-boring-the-determinism-layer",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-07T20:14:49.000Z",
  "source": {
    "name": "Stack Overflow Blog",
    "slug": "stack-overflow-blog",
    "url": "https://stackoverflow.blog/2026/10/07/part-1-make-your-ai-agents-boring-the-determinism-layer/"
  },
  "original_language": "en",
  "account": "The key to safely deploying LLM agents in high-stakes systems isn't about using a more intelligent model, but instead about tightly controlling the agent to operate on a single node. This creates ordinary, testable code around the model, which can be audited and tested. The main principle is to contain the model's intelligence while removing unpredictability.\n\nAn agent functions as a pure function from context to a proposed decision, with the model being the only nondeterministic call. It doesn't have the authority to change the world, and a separate, unchanging component applies proposed decisions after approval. This design turns \"an AI doing something we can't explain\" into \"an AI suggested something, and here is exactly what approved it.\"\n\nThe article walks through a single request through a graph of eight nodes, with only a few being specific to the capability. The model node is the only nondeterministic node, taking structured input, having schema-constrained output, and being retried if there's a mismatch. The verification node is deterministic and contains capability-specific code.\n\nA common issue causing LLM fragility is letting the model return free text which then needs to be parsed. This introduces ambiguity and makes downstream code unstable. A simple fix is to constrain the model to emit a validated structure, treating anything else as a failed call. This approach reduces malformed output and eliminates the need for a parsing layer, providing a built-in guardrail. However, it's important to note that structured output only addresses the parsing failure mode, not the judgment failure mode. Validation is essential at the boundary, ensuring the system only sees well-formed decisions, and the rest of the system can focus on reasoning about the system's behavior.",
  "summary": null,
  "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."
}