首页 分类 维基 关于
SITEMAP 订阅 EN
AI Twitter Hots · 2026-10-28 · 第 047 期 · 10 条 2026 年 10 月 28 日 周三 ·

Cursor ships Design Mode Annotate, Cognition launches FrontierCode 2.0, Anthropic engineer kills coding interviews

Today's highlights

  1. Cursor turns design into agent input: Design Mode adds Annotate — circle elements, sketch on the page, review at different screen sizes — so visual prompts enter the coding-agent loop directly.
  2. Evals move from “meets spec” to “would you merge?”: Cognition ships FrontierCode 2.0; the same day Zed makes Claude Code multiplayer via /delta share.
  3. Human vs agent division of labor keeps rewriting: an Anthropic engineer drops coding interviews; Grok Bot posts both an agent-friendly shipping culture and tibot for mention digests.

1. Cursor Design Mode: Annotate feeds visual prompts to agents

Takeaways: @cursor_ai announces Annotate in Design Mode — give agents visual prompts: edit page elements, draw on the page, and review designs at different screen sizes. Companion shortcut Cmd+Shift+D opens it from the browser. Follow-ups call the “design mode glowup” worth studying.

Why it matters: Coding-agent input is no longer only natural language and diffs — circled annotations on a mock become first-class prompts, and IDE competition expands into whether design and implementation share one session.

Cursor Design Mode Annotate

2. Cognition FrontierCode 2.0: would you actually merge the PR?

Takeaways: @cognition launches FrontierCode 2.0 — most coding evals check whether code meets a spec; FrontierCode asks whether you would merge the PR. Tasks are built by maintainers of the underlying OSS repos (40+ hours each); ~80% of audits yield a reference solution better than the upstream PR. On price/performance: Opus 5.5 and GPT-6 Astra lead the intelligence tier; GPT-6.1 Sol hits cost efficiency at about \$1.03/task; rankings diverge from DeepSWE and the prior FrontierCode release.

Why it matters: Once models can land a green patch, the next bar is maintainability and review taste — benchmarks start simulating real merge decisions, not only unit-test greens.

Cognition FrontierCode 2.0

3. Anthropic engineer: coding interviews are dead on his team

Takeaways: @felixrieseberg (Anthropic) writes that the coding interview is dead for his team — it no longer provides signal, and he removed it from loops this month — plus a spicy optimistic take that junior engineers may gain new openings (linked essay). Bookmarks nearly match likes: a “save this hiring note” style spread.

Why it matters: When a frontier lab publicly drops live algorithm drills, the mismatch between “human solves LeetCode” and “human steers agents to ship systems” is now written into hiring process.

Coding interview dead at Anthropic

4. Grok Bot: tibot watches mentions; the codebase is made agent-friendly

Takeaways: @poteto shares tibot — a Grok Bot that manages an X account: finds posts about you or keywords (tagged or not), daily/hourly digests, spam filter, translation, optional auto-block, and never posts for you. A second post says the Grok Bot team makes the easy thing the right thing so engineers, designers, and PMs land hundreds of PRs a day; token spend largely goes to quality/maintenance, not only new features. Together: a personal intel agent on the product side, and an agent-friendly repo culture on the engineering side.

Why it matters: Personal agents expand from “write my code” to “read my social noise,” and “refactor the repo for agents” becomes a public engineering flex, not only an internal toolchain rumor.

Grok Bot tibot / agent-friendly shipping

5. Claude Code goes multiplayer: /delta share shares thread and uncommitted work

Takeaways: @zeddotdev announces Claude Code is now multiplayer: run /delta share in Claude Code and a teammate joins in their own session with full context and uncommitted work on the same thread, backed by DeltaDB. A follow-up ships Delta CLI for plugging DeltaDB into other tools; next up: DeltaDB in Zed and Pi.

Why it matters: Coding agents move from solo terminal sessions to shareable work threads — the collaboration unit becomes context + working tree, not only a Git branch name.

Claude Code multiplayer /delta share

6. Anthropic: Claude enters drug discovery and life sciences

Takeaways: @AnthropicAI says scientists spent months in deep research partnerships with six organizations developing new medicines, and began drug-discovery efforts of their own; a blog shares early results and how Claude can support stages of drug development. One of the day's highest-engagement official posts.

Why it matters: Vertical “model in the loop” narratives expand from coding and support into regulated life sciences — reliability and auditability requirements will push back on general-purpose agent products.

Anthropic Claude drug discovery

7. trq212: OSS needs a distribution model agents can modify with upstream merges

Takeaways: @trq212 argues for a new way to distribute open-source software so people can modify via agents while still getting upstream merges automatically; coding agents can do this in theory but will not do it automatically. Strong bookmark/discussion energy, rhyming with the day's “agents edit repos / multiplayer sessions” thread.

Why it matters: As the marginal cost of editing code falls, fork-vs-upstream friction becomes a product gap — closer to infrastructure than yet another agent IDE.

OSS distribution for agents

8. thdxr: DeepSeek traffic fits on one rack — what's missing is demand-creating apps

Takeaways: @thdxr puts demand in context: OpenRouter, the most successful model router, can almost serve 100% of its DeepSeek traffic on a single GB300 rack (~\$6–7M of compute); even 10× hidden demand stays under ~\$100M. Companion posts ask where neocloud → inference → router capacity goes without products that create demand — someone's gotta make the apps.

Why it matters: Open-weight hype and GPU buildouts collide with “who builds the apps” — for coding-agent / harness builders, this is a demand-side bell, not a supply-side one.

thdxr OSS inference demand

9. Liquid AI LFM3: ~2B open weights trained for agent harnesses

Takeaways: @omarsar0 highlights Liquid AI LFM3 — a ~2B open-weight family for text, vision, audio, omni, and decision tasks on one backbone. LFM3-1.7B is trained for harnesses like Hermes Agent and Pi with 256K context; Liquid's eval shows 38.4 on WebArena-Verified vs 28.9 for Qwen3.5-9B; fine-tunable on a single GPU — aimed at always-on local agents.

Why it matters: Small-model stories shift from “can chat” to “fits a resident harness” — local agents compete on tool use and long context, not parameter-count ads.

Liquid AI LFM3

10. LangChain: open-sourcing an internal design agent (Content Studio)

Takeaways: @hwchase17 shares how the team built a design agent (with OSS code) for slides, GIFs, images — used multiple times a week; related discussion points to Content Studio, an on-brand Slack design agent. Bookmarks exceed likes: a “take this home” engineering post.

Why it matters: A design agent lands the same day as Cursor Annotate — IDEs absorb design input while growth/brand teams agentify design production; both ends productize together.

LangChain design agent