El Pulso del 25 de octubre de 2026
Las señales que entraron en el radar, ordenadas con sus fuentes y contexto para entender qué cambió.
Claude Duet: Anthropic’s proactive-agent research preview
POR QUÉ ENTRA EN EL RADARDuet is designed to monitor connected apps, create a morning brief, continue work in a VM while the user is away, and request approval before consequential actions. Anthropic explicitly positions credentials isolation, prompt-injection defenses, and an action classifier as part of the product—not add-ons.
ÁNGULO EDITORIAL SUGERIDOAbrir fuente original ↗“The agent that works while you sleep—but stops before it spends money.” Demonstrate the autonomy ladder: read → draft → prepare → ask approval → act.
Decisions API is now GA: fast, bounded decisions for agents
POR QUÉ ENTRA EN EL RADARRather than using a general model for every routing decision, OpenAI offers predicate, choice, and score outputs, claiming up to 10× faster responses than GPT-6 Luna through Responses. It supports up to 400 questions per request, native JSON, images, caching, US/EU residency, and GPT-Live integration.
ÁNGULO EDITORIAL SUGERIDOAbrir fuente original ↗“Stop asking your expensive agent to think about yes/no.” Build a tiny triage demo: retry / escalate / ask user, then explain why a constrained decision endpoint changes agent cost curves.
Kimi K3.1 Pro: 3T-class model, 1M context, open-weight promise
POR QUÉ ENTRA EN EL RADARThe October 23 launch claims K3.1 Pro is Moonshot’s most capable 3T-class model and carries a 1-million-token context window; the announcement says model weights will be released. This is worth tracking as a potential serious open(-weight) option rather than repeating benchmark headlines.
ÁNGULO EDITORIAL SUGERIDOAbrir fuente original ↗“Don’t call it open source yet.” Explain the distinction between API availability, open weights, code, data, and a license—then test it locally when weights actually land.
Voxtral 3: voice-agent stack with sovereign/self-hosted positioning
POR QUÉ ENTRA EN EL RADARMistral is packaging multilingual transcription, expressive TTS, diarization, end-of-turn detection, tool calling, and a real-time SDK together. It supports 50+ STT languages, 35+ TTS languages, and deployment through an EU-hosted API or on a customer’s infrastructure.
ÁNGULO EDITORIAL SUGERIDOAbrir fuente original ↗“Can an open-ish European voice stack replace your call-center agent?” Compare the four hard parts—hearing, reasoning, speaking, acting—and show Spanish/English interruption handling.
Perplexity Computer can run fully local on Apple silicon Macs
POR QUÉ ENTRA EN EL RADARPerplexity says its model, harness, and task execution can now stay on-device, with permissioned access to local files and apps such as Notes, Messages, Mail, and Numbers. It names Qwen 3.6 / PPLX Pro as local options and recommends 36 GB+ memory; cloud escalation remains optional.
ÁNGULO EDITORIAL SUGERIDOAbrir fuente original ↗“Your Mac can now be the agent computer.” Make a privacy-first comparison: fully local vs hybrid vs cloud, and test a harmless workflow on dummy files.
The customer becomes the builder: coding agents and software economics
POR QUÉ ENTRA EN EL RADARThis new paper frames a “software substitution frontier”: functionality customers can build and maintain more cheaply than they can buy. It reports that the 50%-success length of leading-agent tasks has doubled roughly every 129 days, while also stressing that market evidence is mixed and survey responses are not proof of displaced spend.
ÁNGULO EDITORIAL SUGERIDOAbrir fuente original ↗“Will AI agents kill SaaS? The honest answer is narrower.” Use a real example: rebuild one internal feature, then price build + maintenance + failure risk against a subscription.
Creator-watch: Matt Wolfe’s weekly AI roundup (use it as a source map, not the story)
POR QUÉ ENTRA EN EL RADARThe new upload is a useful secondary index of launches: Claude Duet, OpenAI Decisions API, Voxtral 3, Perplexity Computer, Ideogram Vector, and more. The primary links above are the publishable leads; the roundup should not be the main citation.
ÁNGULO EDITORIAL SUGERIDOAbrir fuente original ↗“I traced the AI-news roundup back to the original sources.” Pick three claims, open the original release pages, and label what is announcement, measured result, and marketing.
Creator-watch: Matt Wolfe on Kimi K3.1 Pro
POR QUÉ ENTRA EN EL RADARThis is an early signal that Kimi’s release will quickly reach the creator-news cycle. Publishing an explainer on the release/weights/license distinction ahead of generic “open models caught up” takes would be more differentiated.
ÁNGULO EDITORIAL SUGERIDOAbrir fuente original ↗“Before you download Kimi K3.1: 3 questions nobody puts in the Short.” License, realistic hardware, and reproducible benchmark/task performance.