{
  "id": 15623715,
  "title": "O agente não precisa pensar em voz alta",
  "url": "https://urgent.news/2026/10/19/o-agente-nao-precisa-pensar-em-voz-alta",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-10-19T16:17:11.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/aspepper/o-agente-nao-precisa-pensar-em-voz-alta-pk6"
  },
  "original_language": "pt",
  "account": "Hybrid Reasoning Agents blend latent reasoning with text-based actions, forcing a separation between efficiency, observability, and validation. Unlike LLMs that primarily think and act in text, these agents alternate between latent (unspoken) and text (environmental) modes within a single generation. A lightweight module, HybridLink, manages the transition, while a continuous working memory preserves the hybrid context across turns. During training, two-step processes (HybridSFT and HybridRL) teach the model to coordinate both modes and decide when to remain silent and when to speak. In benchmark tests across 11 tasks and three model scales, Hybrid Reasoning Agents achieved about 10.4% performance improvement over text-based agents and previous latent methods, with token reduction ranging from 35% to 55%, inference speeds between 1.8x and 2.6x faster, and cost savings estimated between 2.7x and 3.7x. For production environments, the key advantage is not merely cost reduction, but the improved efficiency and readability, making it clear when text is necessary (e.g., API calls, diffs, user responses, audit evidence) and when it can remain latent—provided there's a way to recover or sample that internal reasoning when validation demands it. This shift in engineering software assisted by AI changes the design of supervision pipelines. Production teams must measure not just the complete textual log, which could be wasting tokens unseen, but also when to keep the intermediate steps hidden and when to expose them for inspection. The decision to separate when to think in silence from when to expose step-by-step actions is an architectural and governance choice, not a prompt detail.",
  "summary": "Hybrid Reasoning Agents combinam raciocínio latente com ações em texto - e forçam a separar eficiência, observabilidade e validação. Agentes de LLM ainda raciocinam e agem quase só em texto. Cada passo intermediário vira token legível, mesmo quando ninguém precisa ler o raciocínio — só o resultado ou a chamada de ferramenta. O custo sobe. A latência também. Raciocínio latente (em espaço contínuo,…",
  "key_points": [
    "Hybrid Reasoning Agents combine latent reasoning with text-based actions",
    "HybridLink manages transition between latent and text modes",
    "Agents achieve 10.4% performance improvement over text-based agents"
  ],
  "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."
}