{
  "id": 18039555,
  "title": "AI Agents Pick the Wrong Tool More Often as You Add Tools, and Research Shows Why",
  "url": "https://urgent.news/2026/11/03/ai-agents-pick-the-wrong-tool-more-often-as-you-add-tools-and",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-03T22:49:35.000Z",
  "source": {
    "name": "HackerNoon",
    "slug": "hackernoon",
    "url": "https://hackernoon.com/ai-agents-pick-the-wrong-tool-more-often-as-you-add-tools-and-research-shows-why?source=rss"
  },
  "original_language": "en",
  "account": "An AI agent does not autonomously run its tools. It selects a tool from a list, fills in the required arguments, and makes the function call. As the list of tools grows, the model's tool selection accuracy decreases. In a November 2025 engineering post from Anthropic, the company revealed that a setup with five connected MCP servers exposed 58 tools, consuming about 55,000 tokens before the conversation even begins. Anthropic also reported that their internal tool definitions reached 134,000 tokens before optimization. The most frequent failures listed by Anthropic are wrong tool selection and incorrect parameters. When an agent incorrectly calls a tool, such as using search_orders instead of search_invoices, the model still processes the request and generates an answer using the wrong data.\n\nLong prompts are read unevenly by language models. Research conducted in 2024 showed that accuracy drops when relevant information is placed in the middle of a lengthy input, even for models designed for long contexts. A tool list functions similarly to a lengthy input, with each tool definition acting as a separate short document. As the number of tool definitions increases, the likelihood of the model choosing the wrong tool also rises.\n\nIn a May 2025 preprint, researchers Tiantian Gan and Qiyao Sun investigated the impact of tool count on an AI model's selection accuracy. They found that as the number of tool definitions grew, the model's accuracy decreased from 13.62% without retrieval to 43.13% after incorporating a retrieval step, effectively halving the prompt size. Additionally, a 2025 DATE conference paper by Varatheepan Paramanayakam and colleagues demonstrated that reducing the number of available tools to LLMs significantly enhances function-calling performance. This improvement also led to a reduction in execution time by up to 70%, as the model had less text to process.\n\nOverlapping tools present a secondary challenge. Similar tools are harder for models to distinguish, as demonstrated by the MetaTool benchmark presented at ICLR 2024. The benchmark tested models' ability to identify and select tools, especially when the tools were similar to each other. Across eight popular models, MetaTool found that the majority still struggled with effective tool selection.",
  "summary": "More tools make AI agents pick the wrong one. What papers and vendor benchmarks show about tool selection, and how to keep each agent's list short.",
  "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."
}