A Health Permit Filed Before You Noticed, a Pocket-Sized Coder Raised by Agents, and a Perfectly Balanced CI With No Tests Left
Get the next transmission first.
Local preview — subscriptions are not connected yet.
Elementary, My Dear Calendar
Sherlock Holmes never waited to be asked.
He spotted the mud on your boots and told you where you'd been before you'd even sat down.
Anthropic launched Claude Duet on October 21.
It's an agent that works the same way: it comes to you first.
That makes it a proactive agent (one that starts the conversation instead of waiting for a prompt).
What does it notice?
Duet is a research preview for iOS, Android and claude.ai, built on Claude Opus 5.5.
It checks your connected accounts every few hours, sends a morning brief, and only messages you when something needs attention.
In the launch demo, it spots that a new shop opens in 26 days with no health permit filed, so it drafts one.
One tester's call landed right before her flight, so Duet offered to move it.
Boris Cherny just talks his to-do list at it on walks.
Where does it do the work?
Duet gets its own virtual machine with a browser, so our laptops can finally go to sleep.
It reaches over 6,000 connectors and plugins and pays with a virtual card through Link.
Who's holding the magnifying glass?
Before it sends a message, pays, signs or deletes anything, it shows exactly what it's about to do and waits for a yes.
Trading and moving money between accounts are off the table even with approval, and a new classifier watches for leaks of sensitive information.
It can't show us its browser yet, so the deducing happens behind a closed door.
It comes with Pro and Max plans via a waitlist, starting in the US, and it eats into the same usage limits.
The Verge calls it Anthropic's answer to Meta's Muse and OpenAI's Dots.
The game is afoot, and the paperwork is already drafted.
Size Matters Not
Yoda was a small green creature who could still lift an X-wing out of a swamp.
Recursive released Flint on October 21, a coding model small enough to live on a laptop.
The twist is who built it: Recursive says a small team of people and agents made it from scratch in under four months.
How small is small?
Flint is a mixture-of-experts model (only a few specialist sub-networks wake up for each token) with 40B parameters in total, but just 2.5B active at a time.
That's how it manages more than 125 tokens per second on an M4 Pro MacBook Pro.
Recursive says it leads mixture-of-experts models under 50B on SWE-Bench Pro and Terminal-Bench, while running over five times faster than dense models that score the same.
Run it locally, and our code never leaves the laptop.
The weights are on Hugging Face under Apache 2.0.
Who did the actual work?
Recursive's Eureka Machine (its system for automated AI research) ran the architecture search.
It proposes ideas, implements them, tests them and decides what to try next.
It also wrote a Metal megakernel (one GPU program that runs the whole next-token step on Apple chips).
Then it squeezed the output layer down to 4-bit, against common practice, and the model still scores within a point of full precision.
TechCrunch sums it up as a model largely built by AI.
CEO Richard Socher has a new book out called The Eureka Machine.
How do we try it?
Judge it by its size, you should not.
It runs in llama.cpp, Ollama, LM Studio and Recursive's own engine, rec.
Wi-Fi optional, the Force sold separately.
An autonomous on-call engineer that reads your production logs, validates a fix, and emails you what happened… flip on the experimental Auto-republish and it ships the fix too, so the 2am pager finally belongs to someone else.
Ten numbered /presets slots for your favourite prompts, inline /model and /permissions, and image previews right in the terminal… speed-dial for the prompt you've typed 400 times this month.
A busy week of releases adds a claude_code.safety_event for every tool call that gets stopped or prompted, @claude review stop for PRs, and a workspace-trust check on ultrareview… a paper trail for every time the agent got told no.
Run amp orb provider, point it at your dispatch script, and Amp's agents spin up on your own Docker, Kubernetes, E2B or Modal… free for now, with Enterprise pricing landing October 29.
📦 Devin Playbooks Become Skills
Every playbook turns into a skill with the same instructions, and Devin can now auto-review a PR on every push… new accounts from October 24 never even meet a playbook.
Reply to a GitHub notification in Teams, mention @GitHub, and Copilot investigates, runs tests and opens a PR… legacy app users have until November 16 to switch, or the meeting chat stays just a meeting chat.
📦 Copilot Industry Comparisons
The Copilot impact dashboard now ranks your agent-first adoption against peers by company size and industry… opt-in only, which is lucky for whoever's still pasting into ChatGPT.
1.0.100 adds --trust-folder, a telemetry opt-out, and a sandbox that can finally reach localhost, while 1.0.103 carries sandbox settings into ACP sessions… your dev server says hi.
Your coding agent hands Studio the context behind a change, and Studio tests the affected user flows against the live app… because tests written by the agent that wrote the code tend to agree with it.
A read-only server that tells your agent what's valid, running and happening at the gateway, with Datadog, Grafana and Dash0 support… so its config works on the release you actually run.
A US home insurer now lets ChatGPT drive its quote flow through WebMCP instead of clicking around like a lost tourist… your agent can now insure the house it keeps forgetting to lock.
Get on the list.
Local preview — subscriptions are not connected yet.