Welcome, humans.

A new Mac app called Clicky from an indie dev had a demo go viral on X yesterday, and the basic premise is this: it sits next to your cursor, sees your screen, listens when you talk, and can spin up background agents when you say “clicky agent.” As one of the commenters said, the demo had big Steve Jobs energy.

Now, if that looks cool to you, it uses GPT Realtime 2.0 so you can try and vibe-code your own… or Jason Kneen already built an open-source version called OpenClicky, because if there’s a will, there’s an open version (side note: apparently putting the word “open” in your company comes with supposed 10x success multiplier… maybe someone should ACTUALLY make OpenWill, an open source will creator for the docs you need in place before you die…) 

And then there’s Bryce Rattner Keithley, a recruiter who had never written code and still shipped an iPhone app using Replit, Claude, Claude Code, Gemini, Higgsfield, and Kling.

Put those together and you can see the computer interface changing: step one is making computers easier to navigate with voice, vision, and gestures. Step two is making the code underneath easier to create with plain language.

And soon, all the weird little chores of using a computer, like finding the setting, opening the right app, pasting the right command, should all slowly fade into the background of the work (or fun!) you’re ACTUALLY trying to do…

Here’s what happened in AI today:

• 😺 NVIDIA and Microsoft built Windows PCs for personal agents

• 📰 Anthropic confidentially filed paperwork for a future IPO

• 📰 Luma launched an open physical AI research lab

• 🍪 MiniMax M3 brought 1M context to open weights

• 🎓 Codex Goals can work for hours without babysitting

Hey: Want to reach 700,000+ AI-hungry readers? Advertise with us! 

P.S: Remember when we told y’all we would shout out one of your requests from Friday’s newsletter?

Well, Molly from Code.org asked “Would love it if you could shoutout Code.org for the work they're doing building free K-12 education courses on how AI works and how to use it effectively.” Obviously, we had to go with this one! Keep up the good work, y’all!

😺 NVIDIA and Microsoft want your laptop to become an agent machine

DEEP DIVE / FULL BRIEF

Picking up where we left off above, the future computer may feel less like a tool you operate and more like an employee you manage.

That’s the real pitch behind NVIDIA and Microsoft’s RTX Spark, a new Windows PC platform built for personal agents. 

In simple terms: instead of opening apps, hunting through files, and paying cloud credits for every serious task, your computer gets enough local intelligence to do more work on-device.

Here’s what happened:

• RTX Spark PCs will offer up to 1 petaflop of AI performance and 128GB of unified memory.

• NVIDIA says they can run 120B-parameter models locally, which means large models can work on your machine without sending every task to the cloud.

• Microsoft is adding Windows security tools, while NVIDIA is adding OpenShell, a runtime that limits what agents can access and do.

• RTX Spark laptops and desktops are expected this fall from Microsoft Surface, ASUS, Dell, HP, Lenovo, MSI, and others.

How to try it:

• Developers can follow Microsoft Build for the first Windows agent platform demos (Corey’s there; if you see him in person tomorrow, go say hey!).

• Everyone else should watch the fall RTX Spark PC launches.

• The useful test you should run ASAP: can it run a local agent that handles real files, real apps, and real mistakes safely?

Why this matters: Cloud AI turned intelligence into a toll booth. Or, more appropriately, the enterprise equivalent of a video game arcade. 

In my favorite metaphor to describe this possibly ever, Dave Morin of the More or Less Podcast said that everyone at Dell was equating today’s cloud-based AI like video games before home consoles became a thing: everyone has to keep plugging quarters into the machine. Every image, video, code agent, and long task burns credits somewhere. Either you’re paying them or your SaaS app is. 

Local agents change the economics. They make basic computer work feel more like buying a Playstation or a Nintendo game console that you own vs having to get tokens from a Chuck-E-Cheese. The cloud will still matter for top-tier reasoning, giant models, and heavy enterprise tasks. But everyday computer control should eventually become local, private, and built into the machine you already own. This is the first real bet in that direction. 

Well, second*. Apple saw this direction early with Apple Intelligence: the assistant belongs close to your personal context. The problem was timing. The models, memory, and agent safety layer were not ready. But uh, Apple’s WWDC event is next week…  and we’ve come a long way since 2024. Get hyped! 

Our take: NVIDIA and Microsoft are trying to get ahead of the device shift OpenAI and Anthropic may eventually need to face.

If the agent becomes the interface, the local device becomes the main distribution channel for consumer AI. The company that owns the trusted local computer owns the place where everyday AI work happens. And there are already a mighty high number of companies who would rather not pay the troll toll. 

FROM OUR PARTNERS

Knowledge is scattered across all the different apps you use. Data is hard to access for you and your agents. Adapt solves this by connecting all of your tools into one company brain, so anyone — sales, support, ops — can get answers straight from your data in Slack. It even builds the report, moves the 6,000 records, runs the scan, updates HubSpot for you. One company brain, with the brawn to act, for every role in your company.

Test it out with $100 free credits

🎓 AI Skill of the Day: Write a Goal instead of a prompt

Most prompts turn AI into a very polite intern waiting for the next instruction. A Goal turns it into someone you can actually delegate to.

Claire Vo’s Codex Goals walkthrough shows the difference. A prompt says what to do. A Goal defines what success looks like, how to verify it, what cannot break, and when the agent should stop.

That structure let Claire run a Codex task for five hours and 45 minutes, clean 3,900 emails down to 68, and fix hundreds of Sentry errors by having the agent categorize, repair, and replay historical examples.

Use her six-part frame:

1. Outcome: what should be true when done.

2. Verification: how to test it.

3. Constraints: what cannot regress.

4. Boundaries: what tools or files to use.

5. Iteration policy: how to try again.

6. Stopping condition: when to ask for help.

Turn this task into a Goal an AI agent can run without babysitting.

Task: [describe the task]

Write:
1. The outcome that should be true when complete
2. The verification test
3. The constraints that cannot regress
4. The files, tools, or systems the agent may use
5. The iteration policy for trying fixes
6. The stopping condition for asking me to step in

Have a specific skill you want to learn? Request it here. 

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

1. *Less AI theater, more AI acting on your behalf. Dell Pro Max with GB10 gives business owners a serious desktop for prototypes and workflows.

2. MiniMax M3 is an open-weight frontier model with 1M context, multimodal inputs, desktop operation, coding strength, and agent workflows through MiniMax Agent, API docs, and coding setup guides —pricing varies by API/token plan, full open version coming in a few days.

3. Google Gemma released the first iteration of gemma-skills, a repo of agent skills, meaning reusable instruction files agents can load on demand, that helps agents build with Gemma, use MTP (multi-token prediction, a speedup where a model predicts more than one next token at a time), choose the right Gemma model size for a task, and find current Gemma resources.

4. Bloom turns your brand into a system agents can call through API / MCP to make on-brand images, videos, copy, campaigns, and assets from decks, websites, Figma files, and social references —free trial, then $90/mo.

5. LocateAnything finds objects in photos or videos and labels them with boxes or points, with a GitHub repo, HF model, demo to try, and paper —free to try.

6. Typeahead gives you local autocomplete suggestions in every Mac text field, learns your voice, works offline, and keeps your writing on-device —paid once ($79).

7. MolmoAct 2 gives robotics teams open model weights, training data, scripts, evaluation rollouts, action tokenizer, and LeRobot integration for real-world robot-control experiments —free and open-source.

8. NVIDIA Cosmos 3 gives builders open physical AI models for reasoning, world generation, and action generation, with HF checkpoints, GitHub code, datasets, and post-training scripts —free and open-source.

Confused about what an agent is? Watch this!

We’ve been doing a run of beginner-friendly livestreams lately. Our newest, A Total Beginner’s Guide to AI Agents & Automation, covers what agents are, how automation works, and which tasks are worth handing off. Read along here.

Total AI beginner? Start here (goes with this video), then watch the agent one.

📰 Around the Horn

• Zoom launched ZoomMate, an AI teammate that turns workplace conversations into agentic search, custom agents, deliverables, and workflow execution across tools like Salesforce, Jira, Slack, and ServiceNow.

• Anthropic confidentially submitted a draft S-1 to the SEC, giving it the option to pursue an IPO after the review process.

• Luma launched the Open Physical AI Lab, an open science effort focused on world models and generalization for robots.

• Mecka AI raised $60M to train robots from human activity data captured through body sensors and iPhones.

• Straiker found fake Claude Code installer sites stealing developer credentials, AI tool API keys, password vault data, and crypto wallets; eSecurityPlanet covered the campaign too.

• Anthropic’s Mythos reportedly found dozens of critical vulnerabilities at Palo Alto Networks, while raising big questions about scanning costs.

FROM OUR PARTNERS

Your AI is brilliant. It just knows nothing about your company.

The real blocker to AI automation isn’t the models, it’s your domain knowledge. Guru builds the governed knowledge layer your AI is missing.

Correct once. Right everywhere.

→ See Guru in action

🔧 Tuesday Tool Tip: Turn meetings into follow-through

Zoom’s new ZoomMate launch is a useful reminder: the valuable part of meeting AI is not the summary. It is what happens after the summary.

Before your next meeting, write a “follow-through contract” at the top of the agenda:

At the end of this meeting, produce:
1. Decisions made
2. Open questions
3. Owners and deadlines
4. Tools that need updating
5. Draft follow-up message
6. Next action to complete inside [Salesforce / Jira / Slack / Notion / Asana]

That one block tells any AI note-taker, teammate, or agent what “done” actually means. The goal is work that survives the haunted meeting recap that no one wants to actually read through and quickly convert it into the next actionable steps to take.

A Cat’s Commentary

P.S: Love robots? We’re starting a new robotics newsletter! Sign up early here.

P.P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.