Agents

Perplexity Computer is now available in Microsoft Excel, Word, PowerPoint, and Outlook, letting users access Computer from the app sidebar to coordinate work, draft documents, model data, create presentations, and handle email.

The title says a new Siri app preview shows Apple’s plan to compete with ChatGPT. The RSS snippet only discloses an iOS 27 AI overhaul, a redesigned Siri experience, and a standalone Siri app; it does not disclose launch timing or feature parameters.

AutoWebWorld synthesized 29 web environments, 875 pages, and 11,663 verified trajectories at about $0.04 per trajectory, using FSM-defined states, preconditions, and transitions to verify GUI agent tasks instead of human labeling or an LLM judge.

QbitAI says DeepSeek V4 validated Ascend chip-model co-design, with CANN open-sourcing 65 repositories and supporting day-zero adaptation for more than 70 mainstream models, while AIGCode reported 65% MFU in MoE pretraining on Ascend.

NVIDIA Research presented 8 ICRA papers on sim-to-real robotics: ScheduleStream delivered a 3x speedup for multi-arm planning, COMPASS reached about 80% success across 20 real-world navigation trials, and Grasp-MPC achieved about 75% real-robot grasping success.

Mistral upgraded Le Chat into a unified AI agent called Vibe, with one license covering both office work and coding. Existing chats, settings and plans all carry over. Work Mode supports enterprise knowledge search, structured data analysis, document and report generation, scheduled multi-step tasks and reusable skills, and connects to Google Workspace, Outlook, SharePoint, Slack, GitHub and more.

Zhou Zhiwei describes a project-management setup that uses two Git repositories, an AI coding assistant, Shell, and Python to replace at least 80% of manual weekly-update chasing, data moving, chart generation, and engineering-metrics reporting.

Mistral AI announced industrial AI work, a Vibe upgrade, and a 10 MW inference data center in Les Ulis at AI Now Summit 2026; it is working with Airbus, BMW Group, and ASML, and the data center is scheduled to start operating in Q3 2026.

Cognition raised a $1B Series D at a $26B valuation and projects more than $1B ARR by year-end; the post says its valuation rose 2.5× from the $10B Series C eight months earlier, while the rest of the issue summarizes agent, inference, benchmark, and multimodal AI updates from May 26–27, 2026.

Tsinghua THUNLP, ModelBest, OpenBMB, and AI9stars open-sourced PilotDeck; the article says its sub-agent routing reduced cost from $12.58 to $2.83 in a Xiaohongshu content-generation test, while preserving separate WorkSpaces, editable memory, and per-session routing logs.

NTU AutoMan Lab, Harvard, and Xiaomi Auto proposed AutoMoT, a unified VLA driving model using a 4B Qwen3-VL Understanding Expert and a 1.6B Action Expert with asynchronous inference, reaching 89.42 DS and 74.09% SR on Bench2Drive with AutoMoT+.

NVIDIA’s research team open-sourced Polar, an agent reinforcement learning framework that connects GRPO training at the model API boundary without rewriting Codex CLI, Claude Code, Qwen Code, or Pi; on Qwen3.5-4B, Polar raised Codex pass@1 on SWE-Bench Verified from 3.8% to 26.4%, while prefix_merging cut training steps from 1,185 to 218.

xAI released Grok Build 0.1 in public beta through the xAI API for agentic coding tasks, with throughput above 100 tokens per second and pricing at $1 per million input tokens and $2 per million output tokens.

Lemonade CISO Jonathan Jaffe says a single endpoint can run 200 to 10,000 agents, so security teams need to assign identity to each agent and enforce policies at the point of action, beyond current identity and access management systems.

Cognition AI raised over $1 billion at a $26 billion pre-money valuation, while annualized revenue grew from $37 million to about $492 million in one year, and Devin is positioned as an autonomous junior engineer that can plan, test, and deploy through multi-step workflows.

Anthropic describes a six-step Claude Opus workflow for source-code security: threat modeling, sandboxing, vulnerability discovery, validation, triage, and remediation; in its open-source scanning work, it disclosed 1,596 vulnerabilities by May 22, 2026, with 97 already fixed.

Cognition announced over $1 billion in funding at a $26 billion valuation, with enterprise usage up more than 10x this year and annualized revenue reaching $492 million.

OpenAI supports ChatGPT, Codex, and the Responses API connecting to internal MCP servers through outbound-only HTTPS, while teams keep those servers inside private networks.

Anthropic published a zero-trust framework for enterprise autonomous AI agents, saying frontier models compress vulnerability exploitation from months to hours; the post outlines a three-tier architecture, an eight-stage rollout process, and threats including prompt injection, tool poisoning, and memory poisoning.

Meta Platforms is selling consumer subscriptions to Meta AI for the first time, aiming to offset hundreds of billions of dollars in AI investments. The RSS snippet does not disclose pricing, launch timing, markets, or feature differences versus the free chatbot.