Logan has spoken. Gemini 4 Argon is coming.
@saboo_shubham_
-
Task Boundaries Can Control Context Access for Agents With Computers
By
–
Please do!! The context carries – task itself is a boundary (I can choose at run-time what it should have access) and each task is capable of doing high ceiling work because they get their own computer.
-
Agent Audits Need Clear Permission, Identity and Data Responsibility
By
–
if it's my personal agent – the audit will be super clear – my agent took the action with my permission (the transparency that's missing today from personal agents) For cowork or team agents with their own identity, it will be admin + whoever shared the data with them. Just
-
Task Based Architecture Could Move Beyond One Agent Per Machine
By
–
LOL. Task is our primitive, which moves the limit of one-agent one machine.
-

Gemini Agent Gives Each Task Its Own Computer and Model Choice
By
–
Gemini Agent is built for high-ceiling work where EACH task gets its own computer. You can choose Gemini 4 Argon or Claude Opus 5.5. 10 things that make it a universal work Agent (read till the end to see what a REAL work agent looks like): 1. One Agent, one API
Chat, -
EmbeddingGemma 2 enables fully offline multimodal RAG on phones
By
–
What…you can now do multimodal RAG on your phone, fully offline and free.
— Shubham Saboo (@Saboo_Shubham_) 6 octobre 2026
EmbeddingGemma 2 does text search in ~191MB of RAM, or text + image + video + audio in ~567MB. Measured on a Pixel 11 Pro.
Open weights under Apache 2.0 pic.twitter.com/MKvltaGsmRWhat…you can now do multimodal RAG on your phone, fully offline and free. EmbeddingGemma 2 does text search in ~191MB of RAM, or text + image + video + audio in ~567MB. Measured on a Pixel 11 Pro. Open weights under Apache 2.0
-

AI agents share Markdown file as living source of truth
By
–
Markdown is now native in Google Docs and Drive. Hermes, Grokbot, Muse, and Instinct can share one living source of truth: a file I can open, read, and change anytime. The agents write to it. I edit it. The next agent picks up where I left off.
-
2026 personal agents, 2027 collaborative team agents
By
–
2026 is all about personal agents (grok bot, Muse, Instinct, dots). 2027 will be collaborative team agents with identity, security and controls.
-
Free masterclass on building a multimodal voice AI agent
By
–
Just SHIPPED: a free 30 mins masterclass to build a multimodal voice AI Agent that can see, think, talk, and draw at the same time.
— Shubham Saboo (@Saboo_Shubham_) 4 octobre 2026
Go from an problem → idea → spec → working agent as an AI PM. pic.twitter.com/HYBDQU01DnJust SHIPPED: a free 30 mins masterclass to build a multimodal voice AI Agent that can see, think, talk, and draw at the same time. Go from an problem → idea → spec → working agent as an AI PM.
-
Multi-harness RL trains models across several agent harnesses
By
–
Multi-harness RL means training a model with reinforcement learning inside several different agent harnesses at the same time, instead of just one.