首页 分类 维基 关于
SITEMAP 订阅 EN
AI Twitter Hots · 2026-11-04 · 第 054 期 · 10 条 2026 年 11 月 4 日 周三 ·

Claude Code PM Cat Wu departs, Anthropic ships FlashProve, Grok Bot adds calls and shopping

Today's highlights

  1. Claude Code org shock: Product lead Cat Wu says she has left Anthropic; her join story — steering Claude Code to build an RL env without writing a line by hand — is itself a product demo.
  2. Provable inference meets agent surfaces: Anthropic ships FlashProve; Grok Bot pushes phone calls and shopping the same day; Vercel teases its internal security agent @hacker; Devin makes SWE-2 free for a limited window.
  3. Open harness stack keeps specializing: openbot, Beacon, Hermes Cloud, Perplexity Agent Plugins, and DeepSeek Harness Creator cost benchmarks show up together — control planes, memory, hosted runtimes, and eval numbers all competing.

1. Cat Wu leaves Anthropic: Claude Code's product lead signs off

Key points: @_catwu posts that Friday was her last day at Anthropic. Two years ago, when Claude Code was a small internal tool, she asked it to build an RL env — she steered for tens of hours and never manually wrote a line of code, got hooked, and joined. Today the same task can run unsupervised. Chinese-timeline posts from @dotey and others fill in: she led Claude Code / Cowork, previously Scale AI and Dagster, then Index Ventures, joining ~2024-08; last summer she and Boris Cherny briefly left for Cursor and returned within two weeks. No reason or next stop disclosed.

Why it matters: A defining product owner exiting a breakout coding agent is news by itself; the "the product recruited me" origin story also explains early Claude Code better than any changelog.

Cat Wu leaves Anthropic

2. Anthropic FlashProve: provenance proofs for model replies

Key points: @AnthropicAI announces FlashProve — research toward provable inference. Built on established cryptography, it lets developers verify whether a given response came from a given AI model. Coverage frames it as a way for anyone to check model attribution of an AI reply. Engagement is very high (~2.7k likes, ~1k bookmarks).

Why it matters: Once agents place orders, make calls, and touch production code, "did this actually come from that model?" becomes a compliance and supply-chain question. FlashProve pulls the conversation from leaderboards toward auditability.

Anthropic FlashProve

3. Grok Bot: phone calls and shopping heat up the same day

Key points: @elonmusk tells people to try asking Grok @Bot to make calls; the official @bot account ships a new shopping experience — find deals, compare items, track orders. @petergyang posts a real call where Bot negotiates with Comcast (including the quotable "I don't actually have eyes on the sky") and lands a cheaper, faster plan; @poteto summarizes the lifestyle pitch as shop / save / ship with @bot.

Why it matters: Grok Bot's story is expanding from "cloud coworker that codes" into voice calls and purchase decisions. For coding-agent users, whether one subscription covers "call support + edit the repo" changes workflow choice.

Grok Bot shopping and calls

4. Vercel: internal security agent @hacker gets an outward face

Key points: @rauchg says Vercel has an "astonishingly effective" internal security agent nicknamed @hacker that audits every line of code automatically, and that users can now enable similar power on their repos. The same day he plugs npm i -g ai-cli — giving any agent video/voice/image/text generation plus ai decide, all behind AI Gateway cost and ZDR controls.

Why it matters: Platform vendors selling "security agent on every line" signal that the bottleneck after agents write code is trust and review. ai-cli fills in a multimodal tool surface for those agents.

Vercel hacker security agent

5. Devin: SWE-2 free for all subscriptions through Nov 30

Key points: @dabit3 announces SWE-2 is free in Devin (CLI and Desktop) for all subscriptions until 2026-11-30, with a demo video. A clear time-boxed promo to put Cognition's stronger coding tier in more hands.

Why it matters: Limited free access to a stronger coding tier usually produces a burst of real-repo feedback — useful if you are benchmarking Devin against Claude Code / Codex on your own workloads.

Devin SWE-2 free

6. openbot: open-source "Grok Bot alternative" desktop shell (repo not new)

Key points: @Pluvio9yte and others boost nightly-labs/openbot — a local-first desktop app that turns on-machine Codex / Claude Code / Gemini / Grok coding agents into a persistent, collaborating team while reusing existing subscription logins. GitHub shows created_at ~2026-08-12 (already thousands of stars) — today's Chinese-timeline heat is recirculation, not a same-day open-source drop. Security notes still flag broad host access in a development preview.

Why it matters: The demand is clear — keep your subscriptions, add a local collaboration shell. Digests should date the repo so "just launched" framing does not slip through.

openbot

7. Beacon: cross-harness self-improving memory (PH recirculation)

Key points: Posts push open-source Asymptote-Labs/agent-beacon (beacon.sh): share corrections across Claude Code / Codex / Cursor / OpenCode and 20+ harnesses as reusable memory. Repo created_at ~2026-05-12, ~3k stars; today's chatter includes Product Hunt-day amplification — heat recirculation, not a new repository. @liangzhu_AI frames the pain as "you fix it in Claude Code, Codex makes the same mistake in another window."

Why it matters: Once multi-agent / multi-harness is normal, shared memory beats yet another MCP. Still verify local-first claims and default credential/file access.

Beacon memory layer

8. Hermes: Cloud hosting plus v0.22.6 stable

Key points: @witcheer walks through Hermes Cloud — a hosted Hermes Agent on a dedicated instance (own workspace/dashboard), included with new plans; hermes migrate aims at Grok Bot / OpenBot setups. @iamlukethedev announces Hermes v0.22.6 stable: ~776 merged PRs since 0.22.5 across CLI, Microsoft Store, Docker, and Cloud. NousResearch/hermes-agent itself is a long-running, very high-star project (not created today).

Why it matters: Open agents are closing the self-host ↔ one-click cloud gap; migrate messaging aimed at Grok Bot / OpenBot users shows the control-plane fight is now about switching costs.

Hermes Cloud

9. Perplexity Computer: Agent Plugins as an open package format

Key points: @AravSrinivas says Perplexity's agent harness Computer now supports Agent Plugins — packaging app connectors and skills above individual tools. @stromdotcom adds iOS/Mac support and describes Agent Plugins as a vendor-neutral open format bundling Agent Skills and MCP servers; browse a catalog or import from GitHub / zip.

Why it matters: After MCP / Skills, another portable "plugin pack" layer is competing for mindshare. If packs truly move across agents, harness lock-in loosens.

Perplexity Agent Plugins

10. DeepSeek Harness Creator: community benchmarks claim Claude Code parity at ~1/6 cost

Key points: @NFT_Chen and others circulate coding-agent benchmark numbers for DeepSeek Harness Creator: ~63.3% pass rate, matching Claude Code, at ~$3.28 per task (claimed ~1/6 of Claude Code); vs Codex, about 3.4pp lower pass rate (Codex ~66.7%) but cheaper. This is a new eval-angle on yesterday's Creator Mode launch — treat figures as community retellings and chase the original eval write-up before quoting them as official scores.

Why it matters: If an open harness truly holds pass-rate parity with an order-of-magnitude cost edge, default team tooling shifts fast — what you need next is a reproducible script, not screenshot wars.

DeepSeek Harness Creator benchmarks