首页 分类 维基 关于
SITEMAP 订阅 EN
AI Twitter Hots · 2026-10-23 · 第 042 期 · 10 条 2026 年 10 月 23 日 周五 ·

Claude defaults sessions to auto mode, Grok Bot reaches phone apps and 1Password, Claude Models GA on Google Cloud

Today's highlights

  1. Claude defaults to auto mode across clients: Starting next week, Pro/Max/Team sessions on web, desktop, and mobile stop asking for step-by-step approval; automated checks block risky actions. Quoted prior testing: classifiers caught ~89% of dangerous commands vs ~14% with manual approval.
  2. Grok Bot reaches phone apps and 1Password: Calendar, Reminders, Contacts on iOS/Android, plus approved 1Password logins — the personal-agent race heats up again.
  3. Claude GA on Google Cloud, formal verification, and Skills donation: Enterprises can buy Claude with GCP IAM/billing; Frontier Red Team verified 100+ OSS libs with Claude; Agent Skills go to the Agentic AI Foundation.

1. Claude: auto mode by default on web, desktop, and mobile

Key points: @claudeai says that starting next week, sessions across Claude web, desktop, and mobile will run in auto mode by default on Pro, Max, and Team plans: instead of approving each step, an automated check blocks risky actions, and you can switch back to manual anytime (~9.8k likes). The post quotes earlier Claude Code testing where auto-mode classifiers caught about 89% of dangerous commands versus about 14% with manual approval. Community roundups put the change under “fewer interruptions,” now pushed from CLI to every client.

Why it matters: After Claude Code made auto mode the default, the same classifier-backed pattern lands in everyday chat sessions — smoother workflows, rewritten safety boundaries.

Claude auto mode default across clients

2. Grok Bot: native phone apps + 1Password

Key points: @bot shipped two product updates in a row. One: Grok Bot can use Calendar, Reminders, Contacts, and more apps on your phone, rolling out on iOS and Android (~4.7k likes). Earlier: it can use your 1Password account — approve the logins you want to share, and Bot signs in for you over the next few days (~6k likes, 1k+ bookmarks). @elonmusk amplified with “Try Grok @Bot”; @poteto highlighted the native phone-app angle.

Why it matters: Personal agents move another layer down — into system apps and password vaults — as Claude Duet / Muse / Dots keep competing on always-on agents.

Grok Bot phone apps

3. Claude Models on Google Cloud generally available

Key points: @claudeai announced Claude Models on Google Cloud (Anthropic managed) is now generally available: full Claude Platform features on Google Cloud with GCP IAM credentials, billing, and commitment retirement (~2.4k likes). The follow-up says new features and betas ship the same day as on the native Claude Platform — including Opus 5.5 fast mode, code execution, and the Skills API.

Why it matters: Clearer enterprise path — buy Claude against existing GCP commitments without waiting for platform parity.

Claude Models on Google Cloud GA

4. Anthropic: formal verification of OSS libs at scale with Claude

Key points: @AnthropicAI’s Frontier Red Team blog: formal verification can prove code is free of certain bugs, but historically only worked for small codebases or heavy expert time. With @theoremlabs, they used Claude to prove security properties in 100+ real-world open-source libraries over a few weeks (~2.3k likes, nearly 1k bookmarks), and published the research.

Why it matters: AI is not only finding bugs — it is pushing “prove the absence of a class of bugs” to real scale, a counterpoint to yesterday’s PHP2Shell “AI finds RCE” story.

Anthropic formal verification

5. Agent Skills donated to the Agentic AI Foundation

Key points: @bcherny says Agent Skills started as an experiment in Claude Code; after broad industry adoption, Anthropic donated it to @AgenticAIFdn, next to MCP (~1.3k likes). The quoted @ClaudeDevs post is more concrete: since December 2025, 70+ agent platforms have integrated the standard, including ChatGPT, Cursor, GitHub Copilot, and VS Code (~5.6k likes).

Why it matters: Portable “skills” for coding agents move from one product’s experiment to foundation governance — affecting Claude Code and competing harnesses alike.

Agent Skills donation

6. AI tooling security week: Devin’s open-source security plugin, PHP2Shell / LiteLLM aftermath

Key points: @cognition open-sourced Vulnerability Fixes, an agent plugin for recent PHP and Node.js remote-code-execution issues: it helps Devin or any coding agent find affected repos, ask before changing anything, patch to a verified fixed version, test before and after, and open one focused PR per repo (~200 likes). @jeffwang adds that after a big day of PHP / Node / GitLab disclosures and incidents, they open-sourced the plugin and are giving Enterprise users credits to focus on remediation; Devin’s edge is validating that the rest of the app still works after the change. PHP2Shell, the LiteLLM 1.107.0 supply-chain malware, and Railway env-var issues still show up in community briefs, but those launches were yesterday’s main story — noted here only as continuing context.

Why it matters: Once agents hold shells and secrets, patching becomes an agent workflow; open plugins standardize the fix loop.

Devin Vulnerability Fixes plugin

7. T3 Code: smooth on weak networks + noEmDashCopy lint

Key points: @theo says he often didn’t notice degraded internet because T3 Code kept running flawlessly (~370 likes). Separately, he took a copy regression seriously: after someone spotted an em dash, he added a noEmDashCopy lint so it doesn’t happen again (~630 likes).

Why it matters: T3 Code as a Claude Code harness / orchestration layer remains a high-engagement coding product on the timeline; UX details (weak-net smoothness, copy lint) are part of the pitch.

T3 Code noEmDashCopy

8. Gemini 4 Argon: official 3D agent demo vs coding skepticism

Key points: @Google showed Gemini 4 Argon starting from a single Nano Banana image, driving Blender MCP in an autonomous loop inside @antigravity to build a finished audio-reactive 3D cathedral — sculpting Gothic architecture, aligning camera grids, staging dynamic lighting. Meanwhile @NerdSnipePod and others keep asking why anyone would use Argon for coding; the gap between multimodal demos and coding trust is still loud in podcasts and replies.

Why it matters: Multimodal agent demos look strong; coding credibility is not settled — a clean “model vs coding scenario” contrast.

Gemini 4 Argon cathedral demo

9. DeepSeek V4.1 Flash: 100+ tok/s on Mac + OpenClaw

Key points: @NFT_Chen ran DeepSeek-V4.1-Flash-1017 on a single Mac M5 Ultra (80-core GPU + 256GB) via mlx-serve: stable 102–112 tok/s, peak 136.7, prefill ~1770 tok/s, 3-bit quant, wireable from Claude Code / Codex to a local port. @openclaw announced the model is available in OpenClaw and asked for agent feedback.

Why it matters: For open weights in coding / agent land, the questions remain “how fast locally” and “which harness picks it up.”

DeepSeek V4.1 Flash local throughput

10. Reflection Beam + Cline: free 501B and an OpenCode demo

Key points: @cline added new free models including Reflection Beam (501B) and Nemotron 3.5 Super VL (120B), noting that six months ago these Terminal-Bench scores were frontier — open weights make them more accessible. @reflection_ai plugged Beam into @opencode, asked it to study Unsloth and the latest smallest Gemma-4, and build a Text2SQL fine-tuning notebook — out of distribution for Beam — and raised heldout accuracy by about 66.5%.

Why it matters: Open MoEs move from “downloadable” into the default free menu of everyday coding agents (Cline / OpenCode).

Cline free Reflection Beam