AAIF TSC for Google. Core: A2A, A2UI, AI Catalog, ARD, Agent Plugin. Friends with AP2, UCP, MCP + MCP Apps. Muppet+Robot+Dad.
Boulder, CO
Joined April 2008
- Tweets714
- Following367
- Followers318
- Likes2.2K
I work on agents that can work with each other, and with us. A2A for agent collaboration. A2UI for agents driving your design system. Open standards with AAIF. I'll share things I'm building, questions I'm stuck on, and what we learn making these work outside demos.
I've tried "Create a 3b1b style video explainer" for many months and so far it's been pretty terrible, even with skills... I tried to make a grader for "good video" but failed. Do you have a recipe?
You are speaking to my heart here.
You know my background with A/B experimentation, smart routing w/ redundancies, and heroic regression testing as mechanisms to gain confidence, because we don't trust the world... same stuff applies to "untrusted rapidly changing codebases"
Nah... you have a valid ask, but that's not "Eval", instead you're asking for "Package Management" for skills/plugins. We've talked about it. Right now I think everyone just asks the agent to manage.
alan blount retweeted
Just wanted to share a (very much WIP) little side project of mine showcasing the benefits/speed of @googlegemma DiffusionGemma 🙂 (text diffusion model)
Project is called “DiffusionGearvis” (mix between DiffusionGemma and Jarvis, IronMan’s AI assistant) and I wanted to implement some features for which I believe text diffusion can really shine, the first one being computer use.
With the success of Muse from Meta, it is very clear that computer use has tremendous value for users. However, this feature is 1) notoriously slow 2) raises serious privacy concerns (screenshots and other personal data uploaded to the cloud). DiffusionGemma can help with both 🙂 It is super fast thanks to text diffusion and is open-weight so people can run it privately on their own desktop / VM. All the demo videos were run locally on a desktop with a single @NVIDIAAIDev RTX 5090 with speech-to-text and text-to-speech from @kyutai_labs (stt-1b-en_fr and pocket_tts) .
This is with the vanilla NVFP4 checkpoint served on @vllm_project so performance can vastly be improved with dedicated fine-tuning or distillation from larger open-weights models. I hope that in a few months, the open-source community will create some fast “local Muse” on top of it :
Kaggle.com/arenas #1 example is pelicans riding a bicycle... This makes me very happy. @simonw
I had _not_ noticed this before, but now I have seen and understood. Thank you for clarifying why "yet another cli" makes sense. I'll try it.
Agreed. Do we have a public "plan handoff and coordination" benchmark somewhere yet?
Instead of that being an internal, invisible feature, it would make for a great dedicated benchmark IMO.
alan blount retweeted
Muse Spark Open from @Meta is now available on the Gemini platform.
Meta's coding-optimized open-weight model is available for developers building long-running agentic coding workflows.
Try Muse Spark Open now → goo.gle/4zvPTXL
Serendipity. When we planned for Sita to make a UCP / AP2 walkthrough on adventofagents.com/2026/10/1…
... we didn't know it was the day after @google launched the new universal cart in the US. blog.google/products-and-pla…
[Insert A-team "I love it when a plan comes together" meme]
I've been doing this for (checks calendar) um almost 2 years now.
Email is fine, but not quite enough for delegation, planning, and work. We like having a clear set of conventions, a task lifecycle, and a little structure in addition to the text.
Layouts hydrated with values, local logic, and streaming events to sync state. It's about agents driving existing native UI - the meaning rides in the data, not the layout.
Hey do your emails support schemas like booking confirmation metadata? Could we piggyback A2A tasks as email schema and maintain a (slower) task lifecycle over email?