{
  "id": 8990155,
  "title": "Your AI agent is burning tokens on choices that don’t need words",
  "url": "https://urgent.news/2026/09/21/your-ai-agent-is-burning-tokens-on-choices-that-dont-need-words",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-21T19:13:46.000Z",
  "source": {
    "name": "The New Stack",
    "slug": "the-new-stack",
    "url": "https://thenewstack.io/kev-skips-text-generation/"
  },
  "original_language": "en",
  "account": "AI agents are currently consuming a significant amount of computational resources by generating text that is not always necessary. As AI agents make decisions, they send requests to generative models, wait for responses, and parse the output, all while burning through tokens. OpenAI researchers have recently revealed that they spend $7,000 per day running agent workloads. To address this issue, Kev, a new family of open decision models built on Qwen 3.5, offers a different approach by eliminating the generation process entirely. Kev developer Jared Palmer released several versions of Kev, including 0.8 billion, 4 billion, and 9 billion parameter models, on Sunday. Kev operates in a prefill-only manner, processing state, questions, and candidates during a single forward pass before extracting decisions through a pointer head, without resorting to an autoregressive decoding loop. Kev supports three types of decisions: Noul for yes/no, Choice for selecting among candidates, and Score for ordered levels. Developers supply the state and questions, and the pointer head returns probabilities across the available candidates. For instance, in a tool-routing decision, the output may indicate the likelihood of using various tools, such as search (0.82), database (0.13), and calculator (0.05). Kev scores only the candidates provided, preventing it from introducing new options not on the list. Batching decisions can also be performed in a single forward pass, with a block-causal attention mask isolating questions while the pointer head scores each set of candidates independently. While Kev can still make incorrect tool choices, it cannot generate new options. The accuracy of Kev's models, particularly the largest 9B version, reached 83.7% on a locked out-of-domain test, according to Palmer. However, the probabilities Kev returns may not always accurately reflect the developer's confidence in the results. Calibration issues and tradeoffs exist, as evidenced by Kev's performance declines on general-knowledge and arithmetic tests, particularly for smaller models. Despite these limitations, Kev's release under Apache 2.0 provides the advantage of open weights, training code, and evaluation tooling, allowing developers to run and train the model on their own infrastructure.",
  "summary": "AI agents spend a ridiculous amount of compute generating text nobody actually needs. The decisions an agent makes along the The post Your AI agent is burning tokens on choices that don’t need words 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."
}