{
  "id": 14693943,
  "title": "Agentic AI doesn’t have a model problem. It has a retrieval problem.",
  "url": "https://urgent.news/2026/10/15/agentic-ai-doesnt-have-a-model-problem-it-has-a-retrieval-problem",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-15T12:00:00.000Z",
  "source": {
    "name": "The New Stack",
    "slug": "the-new-stack",
    "url": "https://thenewstack.io/agentic-ai-retrieval-problem/"
  },
  "original_language": "en",
  "account": "Agentic AI excels when it has access to accurate and relevant information, but its performance plummets when it relies on poor retrieval systems. Traditional search typically returns a single ranked result per query, but agentic search decomposes tasks into multiple searches, parallelizes them, combines results, and issues follow-up searches as it learns more. This fan-out becomes a normal part of the workflow, especially in search-as-code research, where hundreds or thousands of retrieval operations can occur within minutes.\n\nTraditional search engines return one ranked result per query, while agentic search can fan out multiple searches in parallel, combine results, and issue follow-up queries as it gathers more information. This fan-out increases the workload on the retrieval layer. When one task spawns many searches, retrieval quality becomes crucial. People trust sources they have encountered before and can recognize stale, unreliable, or out-of-place information. An agent must build its understanding from the evidence it retrieves during the task, often starting with little prior knowledge. If an early result is incorrect, subsequent searches can also be wrong, causing the task to drift from its original intent.\n\nTo illustrate the impact of repeated retrieval steps, consider a workflow requiring five retrieval steps, each with a 90% probability of providing the needed evidence. Without retries or recovery, the probability that all five steps succeed is 0.9^5, or approximately 59%. While real agentic systems can retry, use redundant evidence, and branch to recover from errors, this example demonstrates how small per-step weaknesses can accumulate over multiple retrieval steps, especially when later work depends on earlier retrievals.\n\nIn summary, agentic AI's success hinges on effective retrieval systems that provide accurate, up-to-date, and relevant information. A capable model working with poor retrieval results still yields poor material to reason over. As agentic search scales, capacity, tail latency, and retrieval quality become increasingly important to maintain overall system performance.",
  "summary": "On BrowseComp-Plus, GPT-4.1 answered 93.49% of the questions correctly when it was given the supporting documents directly. When an agent The post Agentic AI doesn’t have a model problem. It has a retrieval problem. appeared first on The New Stack .",
  "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."
}