Kagi

I CAN'T F*CKING BELIEVE IT: ANTHROPIC'S ENGINEERS PUT 1 AGENT AGAINST A TEAM OF AGENTS. THE TEAM WON BY 90.2% ON THEIR OWN RESEARCH TEST. i drew it: 1 window fills up, cracks, splits into 4. an illustration, not a live run. you give 1 chat the whole research job. 1 window holds every search. how the team works: → 1 lead agent (Opus 4) reads the question and makes a plan → it starts 3 to 5 sub-agents (Sonnet 4) at the same time → each one searches in its own window → each one hands back only the key parts → the lead puts it all into 1 answer → parallel search cuts research time by up to 90% on complex questions remember this part: it uses about 15x more tokens than a normal chat. 1 agent alone uses about 4x. most people look for a smarter model. on a hard web search test, token use alone explained 80% of the difference in scores. try it today: split 1 big task into 3 small chats. paste only their summaries into a 4th. would you pay 15x the tokens for +90.2%?

Okay this is f*cking wild:I left an old Android charging overnight. By morning it had answered messages, paid a bill, cleaned photos and caught a $482 mistake - all by itself. Ex-Google Android engineer split the agent into 4 roles. One of them only exists to check the other three. What it did: → 3 chats answered → electric bill paid → 3 duplicate photos deleted, 214 backed up → wrong amount caught: $482.00 instead of $48.20 Planner → Reader → Tapper → Typer → Checker. 42 tasks. 19% → 98% battery. This is not sci-fi. Alibaba’s Mobile-Agent-v3 runs the same split (plan → act → check) and scores 73.3 on AndroidWorld. Everyone gives AI a chat window. Almost nobody gives it a whole phone. Your old phone could work the night shift. You’re the one keeping it in a drawer.

5 open-source repos. 286K GitHub stars. One per job to make a video in code. This video used 2 of them: Anime.js and FFmpeg. You keep a folder of 20 tools. You need 5. 01 / Motion: Anime.js A timeline for every move, so motion stays in sync. → github.com/juliangarnier/ani… 02 / Physics: Matter.js Objects fall and bounce like real ones. → github.com/liabru/matter-js 03 / 3D: Three.js 3D scenes in code, right in the browser. → github.com/mrdoob/three.js 04 / Sound: Tone.js Synths and sound effects, so the video is not silent. → github.com/Tonejs/Tone.js 05 / Export: FFmpeg The least flashy one. 65.7K stars. Frames become a file you can post. → github.com/FFmpeg/FFmpeg Stars: Three.js: 115K Anime.js: 72.5K All 5: 286K Today: find the one job your videos miss. Start with that repo. Saving tools feels like progress. Shipping one video is progress.

THIS IS F*CKING WILD. A CHINESE LAB JUST SHIPPED A MODEL THAT SCORES 55. Anthropic's Opus 5.5 scores 58. It costs half. A new model dropped, so I decided to test it. Opus 5.5 vs Kimi K3.1 Pro. 4 rounds, published scores. You pay $20 per 1M output tokens for Opus. Kimi asks $10. The 4 rounds: Code: Opus 66.4, Kimi 65.7. The gap is under 1 point, and Kimi's tokens cost 50% less. Finance: Kimi 64.4, Opus 62.2. So the cheaper model reads reports better. 3D: Opus 96.2, Kimi 92.7. So Opus is ahead on 3D models. Price: Kimi $2 in, Opus $4 in. So every token costs half. Score: 2 : 2. Tiebreak, overall index by Artificial Analysis: Opus 58, Kimi 55. Round scores come from Moonshot's own charts. Moonshot says open weights come by Nov 4. You can try Kimi on its API today. Opus is still smarter. Kimi is close enough to make you do the math. Opus is 3 points smarter. You pay 2x for those 3 points. Your call.

HOLY SH*T. ANTHROPIC LAUNCHED DUET 3 DAYS AGO AND IT WRITES TO YOU FIRST a normal chat waits until you type. duet checks your apps on its own and pings you when something needs a human i went through the launch docs and drew what it does: → plugs into 6,000+ connectors, like your inbox and calendar → checks them every few hours on its own cloud computer, so your laptop stays shut → sends 1 morning brief with only the items that need you → drafts replies and builds dashboards, decks and web apps on opus 5.5 → stops for your yes before it sends, pays, signs or deletes. anthropic's own demo shows a $640 permit fee waiting for approval you wake up and press yes you can't watch its browser. anthropic also says it can misread a thread and miss mail that lands between checks. remember that part no extra price. it runs on your pro or max limits, and background checks count too. research preview, us only, waitlist the video is my sketch of that flow. numbers on screen are demo data no waitlist? run the closest version today. connect your mail and calendar in your ai app, then ask: "be my chief of staff for tomorrow 1. read my inbox and calendar for the next 24 hours 2. pick the 5 items that need me most, most urgent first, and say why for each 3. draft a reply for each one and never send it 4. never send, pay, sign or delete anything. ask me first" start with 1 day of mail and check the 5 items by hand before you trust it would you give it your inbox for a week if it can't send, pay, sign or delete without you?

HOLY SH*T : ANTHROPIC JUST DROPPED DUET. AN AGENT THAT MESSAGES YOU FIRST it runs on opus 5.5 what it does: → sends a morning brief from your email and calendar → checks your apps every few hours and pings you when something needs you → works on its own cloud computer with its own browser, so your laptop can stay shut → builds web apps and slide decks on that computer → finds unpaid invoices in stripe and drafts the reminders before it sends a message, pays, signs or deletes, it asks you first. 4 actions, your ok every time it's a research preview, waitlist only, us only, pro and max plans only. background checks eat your plan limits openai dots starts at $100 a month. duet runs inside your pro or max plan the video is a map of what it can do, built from the announcement which one would you hand over first?

YOU HAVE AN OLD WEBSITE NOBODY DARES TO TOUCH, SO GIVE IT TO ANTHROPIC’S CODING AGENT AND SPLIT THE JOB BETWEEN 4 AGENTS. old code works, so nobody opens it. then one day it breaks, and you have no idea where to start. work on a copy, not the live site. record every page before you change anything. this is your proof of how the site should look. then split the work: → audit: reads the old code and lists what can break, so you know the risks before you fix anything → rebuild: swaps dead libraries for modern ones and keeps the layout, so the site looks the same → data: moves old forms and files to the new version on a copy, so you lose less → test: opens every page in the old and new version and compares them. the agent that fixed a page never checks it give the tester one full path: open the page → fill the form → send → reload → check the data. keep a list of what is blocked or missing. the next agent starts from that list. start with one page. not the whole site. add more only when that one passes the full path. old code is just a pile of risks you can list. agents list them, fix them, and compare the result to the original. the code is not the problem. being afraid to open it is. i test this stuff on myself and share the results. follow