Chris Howard @chrishoward

Actually just downloaded gemma4:12b today. Was working with Antigravity to set up a local server on my older MB Pro M1 Pro to try and take some of the load off of my main local server. Still pretty new to the idea of running models locally in general... (kind of had like no idea up until much more recently than I'd like to admit) So far so good though, and now hearing that QAT keeps tool-calling and structured outputs accurate at 4-bit is super cool/perfect timing. Excited to test this out/keep working with local models. Thanks everybody @googleaidevs

@buildinpublic Building Hattie: an autonomous AI receptionist designed to seamlessly manage calls and book appointments directly to your calendar. We're focused on creating a completely frictionless experience for callers while reducing admin load for the team.

@perplexity_ai That 48x execution time is the real story here. Definitely been seeing this firsthand—people don't want to chat with AI, they want the AI to run in the background and actually book the job.

@NotebookLM Been using NotebookLM daily for research. Adding agentic reasoning takes it from a smart filing cabinet to an actual coworker. Huge update.

Introducing a more powerful NotebookLM 🚀 Massive upgrades deliver agentic capabilities in chat, more advanced reasoning, and a suite of new output formats. Tackling complex, multi-step research problems has never been easier. Rolling out now to Google AI Ultra subscribers.

@PythonDvz 10

@KanikaTolver @Microsoft Right there with you, and the learning seemingly never stops. But I've got to remember how that's a good thing/a blessing

@Nishan_011 At one point I, unwittingly, had 24 or so claude.md's floating around all pointing at each other or randomly.... talk about a nightmare

@googledevs Antigravity's been my most recent daily driver, and i've been really liking the experience. Specifically the ability to pretty easily spin up agents in the cloud to work when I or my laptop aren't "awake". Very smooth when i've prompted antigravity for help with this

Been testing setups to achieve agent architecture's like this one as well as a few others, and lately hosting runtimes on Google Cloud Run specifically. The platform has been highly responsive, though we are still pretty consistently iterating on it. There's no single definitive setup. Having worked with integrations across various models as they evolve, the goal is finding the right coordination to give them "hands", "eyes", and the parts to run autonomously without constant oversight. The core problem I've found being the right balance between developer-defined constraints—manually choosing specific skills and plugins—and letting the agent decide how to coordinate those resources dynamically. I can only remember and then physically invoke so many options when it comes to skills and/or frameworks. The quality of the output varies significantly depending on where you draw that line of autonomy. The main bottleneck is ingesting directory data and routing it correctly into a unified memory layer. Without a structured routing map to reference past runs and coordinate available skills or plugins, the agent loses track. Aligning the memory layer is consistently the most challenging piece. For the perception layer, combining Exa and Firecrawl works quite well, but making sure the playbook is followed every time can also be more of a struggle than you'd think. Making sure the prompts and guardrails within are bulletproof when leaving the agent being prompted to work that flow again and again autonomously is a must, but easier said sometimes than done. Exa handles neural searches to retrieve results based on semantic concepts rather than keyword matching, preventing the agent from chasing stale index pages. Firecrawl then strips Javascript and HTML clutter, converting web pages to clean markdown to save context tokens. Other options like Tavily or Puppeteer exist, but this combination manages the bulk of the input extraction. For hosting, running core agent processes on Google Cloud Run eliminates server management while keeping execution lightweight and responsive. If the agent modifies frontend code, deploying preview URLs via Vercel provides a sandbox to verify visual changes instantly. AWS Lambdas or other platforms are valid alternatives, but the feedback loop between Cloud Run and Vercel previews is exceptionally tight. This layout is just one way of many ways to conceptualize the core coordinates—eyes (perception), hands (connectivity), muscle (execution), and reflexes (telemetry). There's no single right way, just a lot of ways to wire it together. What have you guys been doing lately? What are you using? 👇

@googledevs @googlegemma Directly to your laptop, cool whats the minimum VRAM recommendations to actually run smoothly with agentic workflows?

@OpenAI Very cool. So is it basically a router in front of this that decides which app to hit, so it’s not scanning all 62 every time?

@GeminiApp @GoogleDeepMind @__apf__ @GoogleWorkspace nice, love it

@Schneider6701 @piyush784066 Ive had a great experience since the beginning with them, so yes ? i think is way to say it because my main useage has started over the last couple months or so

I just got @chrishoward from the X Handle Marketplace! Get your own at handles.x.com

@MicrosoftLearn i like this, very cool

Level up your AI skills at #AISkillsFest! Whether you’re in business, tech, leading teams, or just exploring AI, there’s something for you. Get ready for: • Curated playlists for your role • Live digital sessions • Prizes and challenges Register today: msft.it/6016vT2Ca