Modern web guide for agents
It is a Show HN project with 341 points on the front page.
The repo teaches coding agents to use modern browser primitives rather than old JavaScript habits. It is a practical fix for agents writing code from years ago.
Their posts of the last 24 hours, through FxTwitter; plus everclear077/awesome-ai-x-accounts when it answers.
What tech Twitter is discussing, ranked by discussion.
The front page, plus stories over 80 points in the last 24 hours.
Every new issue, each delivered once, looking back up to 7 days.
Anthropic pricingFable 5.2 gets cheaperInput tokens drop from $10 to $8 per million, and output from $50 to $40.» Page 2ProtoMap stops agents from rediscovering dependencies» Page 5Arcjet releases AI security models for coding agents» Page 5Pergent
The tech daily written for one reader
Wednesday, November 11, 2026No. 58Price: $0.04
It is a Show HN project with 341 points on the front page.
The repo teaches coding agents to use modern browser primitives rather than old JavaScript habits. It is a practical fix for agents writing code from years ago.
Upgrade to 8.3.7, 8.2.5, 8.1.7, 7.3.8 or 6.4.6 if your dev server is exposed with --host.

Run agents in one home repo that points to other repos in AGENTS.md, keeping skills and steering files in one place.
Each Expo environment gets its own Supabase backend, and preview builds get their own database.
Plan to update the 26.x, 24.x and 22.x lines shortly after.

Adam Argyle's new technique fixes border radius squish and adds per-item easing and delay.

Input tokens drop from $10 to $8 per million, and output from $50 to $40.
This follows up on Monday's launch. A fast mode costs twice as much and runs 2.5 times quicker. Fable can also act as an advisor that hands easy work to Haiku or Sonnet, or as an orchestrator that delegates to smaller models.
Anthropic

The updates roll out today on web and desktop, and interactive connectors open Adobe, Atlassian and Box inside the chat.
Plan a project on a board, build a tracker in a sheet, or tell a story with your data, all next to the chat. Tools open in full, and you can edit them yourself or ask Claude to edit them with you.
Google AI

On FrontierCode 2.0, Carbon scores 40.0%, up from 26.3% for Argon, its everyday sibling.
Google calls Carbon its strongest frontier model, built for multi-step debugging, long research and hard security work. Argon stays the pick for most coding and office work. It is everywhere today: Cognition added it to Devin, and Perplexity rolled it out to Pro and Max users, saying it costs 2.5x less than Fable 5.2 on agent research tasks.
Anthropic Fable
The prototype boots on a laptop and includes drivers, file storage, windows and an installer.
Anthropic says Fable 5.2 coordinated more than 1,000 subagents for eight days of coding, testing and review, starting from an empty folder. In a separate job, an earlier Fable version moved a service from Python to Rust and made startup about 26 times faster. These are company-reported results, not independent checks.
MiniMax
The H3-Image model has open weights and supports up to 16 reference images for editing.
It runs through MiniMax's API and on Hugging Face, with day-zero support in vLLM-Omni on NVIDIA and AMD GPUs. Launch week has 50% off 16MP-plus generations.

The HN thread has 480 comments on 496 points, a sign people are arguing hard about it.
The post argues that working with AI generates a fragile codebase. Teams shipping AI-written code now carry that risk, and the thread is split on whether it is a tool problem or a process problem.
Dev teams
Matt Pocock's post got 514 likes and 128 replies.
Goodhart's Law says a measure stops being useful once it becomes a target. Matt Pocock argues agents are great at gaming numbers, so counting pull requests as productivity is now actively harmful.
Engineering
Gergely Orosz's post got 2,253 likes and 556 replies.
Waterfall planning failed before because building took too long and plans went stale. If AI cuts build time 100x, planning first may make sense again. Pushback is loud in the replies.
Languages
Guillermo Rauch's post got 583 likes and 107 replies, and Prime's rewrite poll got 197 replies.
Rauch says agents will let purpose-built languages bloom, so there is no one-size-fits-all language. ThePrimeagen polled followers on which language should replace TypeScript's compiler, a test of the same idea.

Export once with torch.export, then the same .pte file runs in Chrome and Safari, fully on your device.
The backend already runs Llama 3.2, Qwen3, Gemma 3, Whisper and others, and supports on-device LoRA fine-tuning. Prompts, photos and audio never leave the tab.
Vercel security
Its agent has already patched 14,665 of them, and 6,029 are rated critical or high.
Guillermo Rauch says he is shocked by how many critical bugs would go unnoticed, and how easily the agent fixes them. You can turn it on for any repo, even if you deploy somewhere else. A separate post says device-bound cookies now cover vercel.com.
Team workflow
His post got 3,349 likes, with machine-aided review before and after each push.
Everyone on his small team is trusted to push to main. They rely on machine-aided review before pushing and human plus machine review after. Pull requests are limited to almost none.
Desktop apps
The team says the port makes launch instant, with no cold start.
Electron is the framework behind many desktop apps. The team is porting OpenCode to Rune, which they say launches fast and that they want to teach others. A big part of Rune was built with Gemini Carbon, Dax says.
Security

An attacker used a login flaw, and GitBook rotated infrastructure credentials and revoked API tokens.
GitBook says customer data was accessible but found no sign that other customers' content was changed. It blames the attack largely on automation and AI, without naming models.
Web platforms

The shutdown post hit 361 points on Hacker News, with 93 comments.
Val Town let people write and host small JavaScript apps in the browser. The files I have do not say why it is closing.
It saves tokens by mapping dependencies so agents don't search for them again.
A family of custom-trained security models for coding agents.
The team says it is up to 100x faster than Parcel 2 on cold builds, and runs without Node.
X is replacing three servers with one app, and a streaming setup fixed slow startup from hundreds of tiny JS files.

Acronis found data theft from universities and emergency services, with the operator copying findings between agents by hand.
HubSpot's CTO argues that ChatGPT and Claude, like browsers, won't own the whole software stack.
He says browsers did not own everything either, and many software companies still got built on top. The shift is back to web-style software.
A post on home automation that drew 867 points and 309 comments on HN.