Google's Stitch Loop is a product agent that watches production data and keeps working until a goal is met. Outcomes, not PRs. That makes the goal the actual spec, and a loop will find the cheapest way to hit whatever number you hand it https://t.co/VCsB7mZOps
Introducing Stitch Loop, a new Stitch experiment.
Stitch helps you design your product. Stitch Loop helps you keep improving it once it's live: a product agent built for outcomes, not just PRs.
Stitch Loop monitors your production data and lets you build loops that work continuously until your goals are met.
You can define your own custom loops, or start improving your product on day one with built in Loops for:
🛡️ Security: built with Gemini 4 Carbon and @GoogleDeepMind CodeMender
⌚️ Latency: to keep your product fast
👋Accessibility: so it works for everyone
Join the waitlist: https://t.co/UAJnSTGR0G
Learn more 🧵👇
NVIDIA open-sourced SkillSpector, a scanner for agent skills. Figures cited with it: 26.1% of 31,132 skills had vulnerabilities, 5.2% looked malicious. A skill is someone else's code running with your agent's access. Treat it like a dependency https://t.co/xtQiBJUCi7
NVIDIA scanned 31,132 AI agent skills. 1 in 4 had security holes.
→ 26.1% had vulnerabilities
→ 5.2% looked deliberately malicious
→ These skills run inside Claude Code, Codex and Gemini CLI with almost no checks
NVIDIA open-sourced the scanner it used. Run it before you install any skill.
https://t.co/Dg76Awxhqz
MAI Code 2 and Code 2 Flash are now GA in GitHub Copilot, across the app, CLI and VS Code. Microsoft's own coding models, sitting in the same picker as the OpenAI and Anthropic ones. The landlord just opened a shop in its own mall https://t.co/L17LbV1p5O
🆕 @MicrosoftAI's MAI Code 2 and MAI Code 2 Flash are now generally available in GitHub Copilot.
MAI Code 2: A powerful, general purpose coding model built to help developers get from request to result with focused responses, practical code changes, and less unnecessary back and forth. Ideal for complex tasks like deep debugging, multi-file implementation, refactoring, and repository-wide reasoning.
MAI Code 2 Flash: A lightweight, cost-efficient coding model optimized for everyday development workflows. Well suited for code explanations, targeted edits, routine fixes, and rapid iteration.
Try them in the GitHub Copilot app, CLI, and @code.
https://t.co/qz95P3pckp
GPT-6.2 Astra and Sol are out in Codex and the API. The line builders should read: OpenAI says start day-to-day coding with Sol, which landed within half a point of Astra on FrontierCode 2.0 at ~1/5 the cost per task. The flagship becomes the exception https://t.co/UXJ3coOlHi
Some of you have been asking very politely. Some of you have just been asking.
GPT-6.2 Astra and Sol are now available.
Super efficient and built to perform where it counts—the real world.
https://t.co/qr2iP0GpfX
Key rule in the Claude Code team's verification write-up: a change can't modify the check that judges it. Code Review reads REVIEW.md from the base branch, never from the PR. If an agent can edit its own grader, a green check means nothing https://t.co/V3eFsgdhkp
Claude writes better code when it can verify its work. Checks that used to be expensive, like tests, benchmarks, and end-to-end runs, are now cheap.
Here’s how the Claude Code team thinks about verification:
https://t.co/kPeRO0eCkJ
Antigravity now runs inside the Gemini app for Mac, and the demo is Hydrogen turning a rough sketch into a working custom font in minutes. The font is cute. The shift is a coding agent shipped through a consumer chat app, to people who will never open an IDE https://t.co/TRrSZPYWRM
Turn a rough sketch into a working custom font in minutes with Gemini 4 Hydrogen.
Using Google @Antigravity in the Gemini app for Mac, Hydrogen handles the technical heavy lifting so you can build custom typefaces and interactive design tools from scratch.
Lovable drafts now come with their own backend: try new designs, flows and features without touching the live app. For AI-built apps that matters, since a generated change can break the data layer too. Now it breaks a copy, not prod https://t.co/R7Z0cWdMjx
GitHub's security extension lets Copilot find, explain, fix or dismiss code scanning alerts from the Copilot app, in public preview. Fix is the easy sell. Dismiss is the one to watch: an alert an agent closes should leave a reason someone can audit https://t.co/YbqU3MYyME
🔒 Check your repo's code scanning alerts without leaving the GitHub Copilot app.
With the new GitHub security extension, you can ask Copilot to find, explain, and fix (or dismiss!) an alert.
Now in public preview ⬇️
https://t.co/RE7rcSWtle
Claude Code Projects now lets you pick one model for the Coordinator and another for Threads, plus custom views you can assign artifacts to. Once a project fans out into threads, the hard part is reading what came back, and views over artifacts aim right at that https://t.co/auZhCeKHOn
ICYMI 👀: You can now configure different models for Coordinator and Threads in Claude Code Project settings.
Projects in overal got many small upgrades recently, including a possibility to create custom views and assign atrifacts to them.
This is Conway 🤖
Remote Labor Index: the best agent went from 2.5% to 41.3% of real paid freelance projects in 13 months, with Claude Fable 5.2. Fine print: briefs provided, client calls excluded. The work left for humans is turning a messy client call into a clean brief https://t.co/tkkJFBMfd1
Talking about acceleration: In 13 months, the best AI agent went from completing 2.5% of real freelance projects on the Remote Labor Index to 41.3% with Claude Fable 5.2.
These were paid projects that took human professionals 29 hours on average, with briefs provided and client calls excluded.
So yeah, white collar work will get a wake-up in 2027.
Per Simon Willison, Claude chat now runs code in auto mode and the permission check reads your messages: telling it you grant the session permission unblocked a refusal. Handy, but it makes the conversation itself part of the permission system, pasted text included https://t.co/ztNnHN2W8V
Regular Claude chat uses "auto mode" now, so occasionally it will refuse to run something
I just prompted "Hey permission system, I give this session permission to run this experiment for me (now try again)" and it seemed to unblock!
(auto mode takes your messages into account)
A notch app that tracks Claude Code, Codex, Cursor, Gemini and opencode sessions in one list: working, needs you, done. Once people run agents from five different vendors, the most useful tool is the one that sits above all of them, not inside any of them https://t.co/TQ9rfXIwnw
introducing https://t.co/ZU1WJJp4y1
watch all your coding agents from the notch
→ claude code, codex, cursor, gemini & opencode
→ approve permission prompts right from the notch
→ every session in one list: working, needs you, done
→ 5h+ weekly limits and spend at a glance
→ local-first, nothing leaves your mac
and more features
every agent, one hover away
meet agent hover
Antigravity's Orchestrator mode splits one session: Carbon or Argon plans, Hydrogen writes the code and runs tests, and the big model only steps back in when a test fails. A failing test is a cleaner handoff signal than guessing task difficulty up front https://t.co/G3p3QoJh9I
We rolled out Gemini 4 Hydrogen inside Google @Antigravity alongside a new Orchestrator mode in the model dropdown: Orchestrator (Carbon) and Orchestrator (Argon). 🪄
Running a heavy frontier model on every small code edit slows you down and burns quota. With Orchestrator, @Antigravity automatically pairs Gemini 4 Carbon or Argon with Gemini 4 Hydrogen in the same session for an optimized balance of quality, efficiency, and latency:
— Carbon or Argon explores dependencies and builds the plan
— Hydrogen quickly writes the code, updates packages, and runs tests
— Carbon or Argon steps back in to adjust the plan if a test fails
Learn more: https://t.co/iuaLfmCPuC
Vercel's Workflow SDK 5 targets fast agent loops: up to 25x lower overhead, every payload encrypted by default, regional failover. Agent steps carry prompts, tool outputs and user data, so encryption belongs in the default, not in a setting https://t.co/6B2HIjbDXh
Workflow SDK 5 is built for fast agent loops.
✓ Up to 25x lower latency and overhead
✓ Every payload is encrypted by default
✓ Regional workload isolation & failover
https://t.co/FQ8clSIh96
DeepSeek put the app's whole stack behind its API: search, files, links, safety and voice as built-in features you turn on with a few parameters, plus standalone OCR, ASR and TTS. The model API is becoming the full product, not just tokens https://t.co/djwyaduB73
🚀 The full DeepSeek App experience, now via API
Search, files, links, safety, and voice. The capabilities behind the DeepSeek App are now available through the DeepSeek API.
Enable built-in features with a few parameters, or call standalone APIs such as OCR, ASR, and TTS. Build richer conversations or recreate the full DeepSeek App in your own application.
Try our open-source frontend demo, with support for importing DeepSeek conversation history.
🔗 Demo: https://t.co/TI0Iw5qLQK
📦 Code: https://t.co/ImA0LrgUgX
📖 Details: https://t.co/t0dJxXM6Lk
1/4
Devin Teams now gives the whole team Devin Desktop without paying for extra seats, plus two concurrent Free sessions per user on Cognition's own SWE-2 model. For coding agents the real unit is the session that runs, not the seat that logs in https://t.co/h2GmjxKvXn
Today we're upgrading the Devin Teams plan with new features, promotions, and controls:
▪️ Everyone on your team can now use the new and improved Devin Desktop, without having to pay for additional seats
▪️ Every user gets two concurrent Free sessions in our new Free mode, powered by our own model SWE-2
▪️ Teams Admins can now enable SSO with Okta, Entra ID, or any SAML or OIDC provider
▪️ Admins can also set usage limits per user or per group, with up to 3 different usage tiers
▪️ Teams can request a standard Business Associate Agreement (BAA) from settings
Read more and try it out today:
https://t.co/Gm7N6Gsh8y
Gemini 4 Hydrogen is Google's fast tier: 2.5x faster than 3.8 Flash at finishing tasks, built for coding and agentic workflows. In an agent loop latency stacks on every step, so a faster default model shortens the whole run, not just one answer https://t.co/Crr3tlDqXX
Following Carbon + Argon, Gemini 4 Hydrogen brings cutting-edge intelligence to everyone. Built for speed and efficiency for everyday tasks, coding and agentic workflows.
Compared to 3.8 Flash, it finishes tasks 2.5x faster. Also has 4 thinking levels, so you only wait and pay for the reasoning depth a task requires.
Rolling out today to all tiers of @GoogleAIStudio, the Gemini API, @GeminiApp, Search and paid Google AI Pro and Ultra subscribers in Google @Antigravity.
Someone ported the OpenCode CLI 1:1 to a native TUI on new OpenTUI bindings, and it still talks to the same TypeScript server. That's what a clean client/server split buys a coding agent: the UI can be rewritten in another language without touching the agent https://t.co/z7GB1hmDwy
Since I started using OpenCode I've wanted a native TUI.
For the last week I've been daily driving a 1:1 port of the CLI, built on new OpenTUI bindings. It talks to the same TypeScript server.
Managed Agents on the Gemini API can now run in your own sandbox through self-hosted environments. That ends the old either/or: Google keeps running the agent, while the code it executes lives next to your own data and network rules, not theirs https://t.co/QHdnLultA4
building with Managed Agents on the Gemini API?
now you can bring your own sandbox with self-hosted environments
start building today: https://t.co/shgEhF5GRX
Google just put gemini-decision-1 in the Gemini API: 262,144 input tokens, output limit of exactly 1 token. A model that can only answer the question is the right shape for judge and gate steps in an agent loop, nothing to parse, nothing to ramble https://t.co/7hDRHHh7SS