Mr. Ånand

Damn 💀

Hermes Agent on GPT-6 Luna used more of its feedback budget on ARC puzzles it had already seen. 20% the first time, 29% when the puzzle came back

budget here means paid demos and wrong answers, and a fail burns all of it. whatever it kept between puzzles didn't make the repeats cheaper

a plain baseline that saves one working program per puzzle and runs it first on return drops to 10%

𝗖𝗹𝗮𝘂𝗱𝗲 𝗰𝗮𝗻 𝗻𝗼𝘄 𝘁𝗮𝗸𝗲 𝗮 𝗺𝗲𝘀𝘀𝘆 𝗽𝗹𝗮𝗻 𝗼𝗻𝘁𝗼 𝗮 𝘁𝗲𝗮𝗺 𝘄𝗵𝗶𝘁𝗲𝗯𝗼𝗮𝗿𝗱, 𝘁𝗵𝗲𝗻 𝗰𝗮𝗿𝗿𝘆 𝘁𝗵𝗮𝘁 𝘀𝗮𝗺𝗲 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗶𝗻𝘁𝗼 𝗮 𝘀𝗽𝗿𝗲𝗮𝗱𝘀𝗵𝗲𝗲𝘁 𝘄𝗶𝘁𝗵 𝗳𝗼𝗿𝗺𝘂𝗹𝗮𝘀 𝗮𝗻𝗱 𝗺𝘂𝗹𝘁𝗶𝗽𝗹𝗲 𝘁𝗮𝗯𝘀, 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝘆𝗼𝘂 𝗿𝗲𝘀𝘁𝗮𝘁𝗶𝗻𝗴 𝘁𝗵𝗲 𝗽𝗹𝗮𝗻.

Three pieces, one shared context:

→ Whiteboards: the team lays out and organizes ideas together in a meeting, then asks Claude to summarize the findings

→ Sheets: the team and Claude build the spreadsheet together, organize the data, edit formulas and work across multiple tabs. You can also ask for custom artifacts like an account tracker, a daily briefing, a project status board or a contract review table

→ Data (formerly Dashboards): carries the same context forward without repeating details. Connects to tools like Snowflake and Salesforce, uses definitions your data team certifies, and builds dashboards, visual data stories or narrated short videos

Interactive connectors embed the full app inside Claude, so you can work in Adobe, Atlassian or Box beside the conversation. Change things yourself or ask Claude to do it with you. Claude Marketplace lists 9,000+ connectors and plugins, including community-built ones.

Early numbers from Anthropic:

→ Its legal team built a Slack review tool in Claude, and marketing reviews went from 3 days to 24 hours
→ Advantage Solutions cut FP&A forecast assembly from about 10 hours a cycle to under 30 minutes

A coding agent with full permissions on your laptop can read your API tokens, rewrite your .git and, with a browser tool, act as you on any site you're logged into.

Three new open-source tools each fence off a different part of that:

Keep it off your machine
1. devvm: one Lima VM per project folder. It mounts only that folder, new VMs reach only the hosts you allow, and your GitHub and Claude Code tokens stay on your machine. With devvm --clone the VM gets its own clone, so you git fetch its work and review it before merging.
github.com/devvmhq/devvm

Hand the risky step back to you
2. browse: an MCP server and Chromium extension that lets agents drive background tabs in your real browser. At a login, a 2FA prompt or a payment, it hands the tab back and waits for you to click Continue. Clearing cookies or closing the browser is always refused.
github.com/anomalyco/brow…

Ask before the irreversible stuff
3. Agent Autoloop: runs Claude Code or Codex in the background from schedules, GitHub issues or alerts. Before deleting data, restarting a service, spending money or messaging people, it sends you a card on Slack, Discord, desktop and more, and no agent process runs while it waits.
github.com/KKKKhazix/agen…

All three are MIT licensed

𝗤𝗼𝗱𝘂𝗰𝗸 𝗵𝗮𝘀 𝗮 𝗻𝗮𝗺𝗲 𝗻𝗼𝘄: 𝗤𝘄𝗲𝗻𝟰-𝗠𝗮𝘅-𝗣𝗿𝗲𝘃𝗶𝗲𝘄

On Saturday it was a mystery model in Qoder's picker. Turns out it was a Qwen test build, updated daily for 4 days.

What changed since then:
→ Name: Qwen4-Max-Preview
→ Access: every Qoder product and plan, Teams and Enterprise included. On Saturday the docs left out the IDE, JetBrains plugin and CLI
→ Credits: 0.1x for a limited time

Qoder says today's build is newer than the one people tried this week