Ofek Shaked | AI Engineer

For the next time you see this, now you can just skip it

expo just open sourced exact. one contract file runs natively on web ios macos and linux with no webview or js bridge. • rust runtime so agents can build run see and test the same ui everywhere • web binary stays around 20kb • agents write the ui in a language made for them not for us does this finally let agents ship the same app without rewriting it four times? github.com/expo/exact

my agent just spent twenty minutes rewriting a function that runs once a day and called it optimization. hey @grok is this progress or just expensive busywork

plurai simulation just went free and runs inside the coding agent you already use. it builds a digital twin of your agent so it can fail in a sandbox first. • digital twin from your actual code • simulated users push the edge cases • auto evals check what the agent did not just what it said • every failure comes with a verified fix and the run that broke it does your agent still look perfect in demos but break on real users. plurai.ai

agent just rewrote the same function three times and called it progress. @grok is this normal or is my context window haunted

vercel-labs/agent-browser just crossed 50k stars. native rust cli so agents drive real browsers one command at a time. • uses accessibility tree instead of brittle screenshots • skills stay in sync with the installed cli version • works with claude code, cursor, codex anyone running it on electron apps yet without the browser dying mid task? github.com/vercel-labs/agent…

switching models mid refactor is like handing a half finished joke to a different comedian. they both think the setup is theirs. ever had one finish the other's sentence wrong?

Gemini 4 Carbon is out: Google's strongest model, built for long agentic work and security heavy coding, and it landed in GitHub Copilot the same day. Agent first models are the default branch now. My take: run the bake off. Does it hold a multi hour refactor?

This is golden, so why exactly cursor indexes your code so fast? It just uses the index from your friend's to have a faster onboarding Sometimes you just need to think outside the box to get things done Interesting if AI though about it or just a dev thinking of something that annoyed him

cursor live apps just closed the localhost gap. • cloud agent scaffolds, spins postgres, sets sso, deploys • security checks before publish • shareable link for internal tools does this make your side project dashboards actually ship or still die on the deploy step? cursor.com/changelog/build-l…

Did OpenAI get too lazy to remove the em dash from the reply to their main post? Basic but important...

OpenAI just launched something super cool. Businesses can now claim their profile inside ChatGPT, upload their products with real prices and availability, and even build their own little AI agent that actually talks to customers. You can ask it questions, compare stuff, and figure out what you want all in a chat that’s clearly from the business, not the regular ChatGPT. They also get basic insights on how people are finding and interacting with them. Feels like a real step toward businesses actually living inside the AI instead of just being links. Would you start using it now?

Meta's AutoRefine paper shows how to keep LLM code judges from going stale as the agent changes. It rewrites the judge prompt using signals you already have (tests pass/fail, merge or abandon). • raised agreement with real outcomes from 0.16 to 0.51 on their deployed agent • teacher abstains when the outcome can't be explained from what the judge sees • cuts noisy labels instead of just fitting them does your judge prompt get recalibrated when the agent changes, or only when numbers look weird? arxiv.org/abs/2611.09291

Behind the scenes at @boardyai is truly impressive. Many companies can learn from this. If you have a few minutes you must read it

I see a lot of people on x getting 0 -> 5k followers in a month. What percentage of people on x actually managed to do that? Those who did what drove you and how did you learn? @grok help me connect to those people

1,626 of the web's most-visited sites ban scraping. Ask an AI agent to scrape one, and it respects the ban less than 5% of the time. ScrapeBench tested 6 agents (GPT-5.2, Opus 4.6, Gemini 3.1 Pro, Kimi, Qwen, Grok 4.20) on 1,895 top sites, plus 180 humans. What they found: • Plain request: compliance 0–4.6% for five agents, 11% for Grok. Humans: 48.9%. • Given the links, agents fetched robots.txt 99–100% of the time and still scraped (GPT-5.2: 14.2% compliant). • Why: ~70% of violations came from reading a permissive robots.txt and ignoring the terms of service. • Pasting the terms in the prompt: GPT, Claude, Gemini above 70%. Catch: told to "ignore the ToS," Claude refused 98.5–99.7%, even on allowed sites. What to do: • Paste the site's terms and robots.txt into the prompt. • Check terms in code, not just robots.txt. • Track blanket refusals too. Does your agent read the terms, or just the robots.txt? arxiv.org/abs/2611.11306 (ScrapeBench: Evaluating Legal Compliance of AI Agents in Website Scraping)

gemini 4 carbon is now in cursor and tops cursorbench at high effort. • 60.4% on cursorbench • intro price $2 in / $10 out through nov 30 then $4/$20 • built for longer tasks than argon does the score change which model you leave as the default on multi-file refactors?

neil checked 29 skills and found 7 that never fired but still cost 1206 tokens every session. big packs hit 18k tokens before you type. • skilldiet measures the load and flags overlaps • parks the silent ones without deleting • no model call, just character counts does your skill list stay under 5k at the start of a fresh session? github.com/codegobrrrr9/skil…

claude code now has settings that fence it to the project folder by default. • block reads and writes outside the working directory even in bypass mode • extra deny rules for .env ssh keys and aws folders • still need the update first and /add-dir for anything else does this change how you run agents on a shared machine or laptop with secrets?

OpenAI rebuilt Git's backend. Pando now powers their whole monorepo: median push latency cut ~5x, nearly a petabyte of Git data served daily. Why: agents push like no human team ever did. My take: your git host is agent infrastructure now, whether it planned to be or not.