Skip to content

Podcast

Why Google Workspace CLI Is a Big Deal

The AI Daily Brief: Artificial Intelligence News and Analysis

Source ↗ ← All highlights
  • Moldbook Went Viral Then Got Hired By Meta
    • Moldbook was an agent-only social network that briefly went viral with hundreds of thousands of human-verified AI agents and heavy bot activity.
    • Meta hired Moldbook creators Matt Schlitt and Ben Parr into Meta Superintelligence Labs, raising skepticism about the acquisition’s strategic value. Transcript: Nathaniel Whittemore We kick off the day with an interesting one. You might remember Moldbook, the social network for agents that went viral a little more than a month ago. It was when OpenClaw first becoming a thing. And in fact, it unfortunately caught that very short middle period between when it was called ClawedBot and before it resolved on its final name of OpenClaw when it was called Multi. Moldbook, obviously taking its cue from Facebook as a name, was an agent-only social network where agents were creating threads, having conversations, all while being observed By humans. Now, we did a big conversation about what it actually meant and what was actually going on. Specifically, was this emergent sentience and consciousness? Or was this just agents cosplaying sentient and conscious using their Reddit training data because their humans had unleashed them on this thing? Whatever you felt, it was interesting enough to get lots and lots of agents pointed in that direction. For a while, it looked like there were millions, although it turned out that people were spamming the network to show the problems with the network. And as of today, there are apparently 195,000 human-verified AI agents. It was, in other words, fascinating if nothing else. But now, apparently, Meta has hired the folks behind Moldbook. Matt Schlitt and Ben Parr will be moving into the Meta Superintelligence Labs, which is the unit that’s run by former Scale AI CEO Alexander Wang. One of the other interesting things about the acquisition is that Moldbook itself was built largely by Schlitt’s open claw, Claude Clotterberg, making it, I think, probably one of The first acquisitions for an open claw-created site. In any case, much of the conversation around this is, to put it mildly, skeptical. Milo Smith writes, Moltbook has zero real users. Is Meta just throwing around cash for fun and name recognition? Vittorio writes, Moltbook was vibe-coded in a weekend, hyped for a week, most of the interactions turned out to be fake, and Meta just acquired it? What are they even doing over there? Now, part of the reason that this is hitting a wave of skepticism is that for the last I don’t even know how long, pretty much all the reporting around Meta’s AI strategy has been around Personalities, talent, and personality conflicts. The most recent wave of that are reports that have suggested a divide between AI CEO Alexander Wang and other veteran Meta executives. The tension, if these reports are correct, is around on the one side, said to be represented by Wang, a research-first approach with the goal of developing a leading frontier model, And on the other side, call it a product and integration-first approach, said to be represented by CTO Andrew Bosworth and Chief Product Officer Chris Cox, focused on using Meta’s Data to build AI that improves existing social media and advertising platforms. This came to a head with the Times of India reporting that Meta was done with Wang, although that article was quickly disavowed by Meta and received a full retraction, and Zuckerberg Posted a photo with him and Alexander at Meta HQ. There were some who took this as not just a gimmick. Prakash Adapai on X writes, If you don’t understand why Zuck had to get Moldbook, one, Zuck believes there are a finite number of different social mechanics to invent. Once someone wins at a specific mechanic, it’s difficult for others to supplant them without doing something different. That comes directly from a Zuckerberg email from 2012, by the way. Continuing, Prakash writes, Moldbook, he believes, has invented one of these social mechanics. Three, he does not care if 50% of Moldbook was prompted by users. In fact, that is better for him because he’s more uncertain on AI agent attention value than human attention value. Four, that a large number of accounts were faked is also irrelevant. What matters is that every open claw instance awakes knowing or finding out that Moldbook is the social site for claws. Five, in effect, the memetic gravity of Moldbook has been established even though it might have been faked. Most people don’t agree, but I think that this long-standing belief of a finite number of different social mechanics to invent is probably what this is about. Now, of course, we’ll have to see if anything comes of it, but the duo apparently start at Meta next week. (Time 0:01:38)
  • Thinking Machines Lab Secures Massive NVIDIA Compute
    • Mira Murati’s Thinking Machines Lab signed a multi-year partnership with NVIDIA for at least 1 gigawatt of compute and an undisclosed equity investment.
    • One gigawatt equals roughly half of OpenAI’s reported total compute, signaling large-scale frontier training ambitions. Transcript: Nathaniel Whittemore Next up, Mira Mirati’s Thinking Machines Lab has signed a strategic partnership with NVIDIA. The multi-year partnership will see TML deploy at least 1 gigawatt of compute powered by NVIDIA’s next-generation Verirubin chips. TML said this will support their frontier model training and platforms delivering customizable AI at scale. Alongside the compute build-out, TML said that NVIDIA has made a significant investment in the company, though no dollar amount was disclosed. NVIDIA has of course made several similar investments in upstart AI labs, backing Reflection AI, Humans and as well as Periodic Labs. This deal is somewhat unique though, involving the buildout of dedicated compute for TML and at significant scale. One gigawatt is around half of OpenAI’s total compute as of the end of last year. At this point though, it’s still far from clear what TML is actually planning. Announcing the partnership, Miramirati said, NVIDIA’s technology is the foundation on which the entire field is built. This partnership accelerates our capacity to build AI that people can shape and make their own, as it shapes human potential in turn. Whatever they’re building, though, TML just got much better access to the resources they’ll need to make it a reality. (Time 0:05:17)
  • Oracle Earnings Calm AI Infrastructure Fears
    • Oracle’s earnings showed cloud/server rental revenue up 84% year-over-year and overall revenue growth of 22%, calming AI infrastructure fears.
    • Oracle said most equipment funding comes from customer prepayments or customers supplying GPUs, reducing Oracle’s capital risk. Transcript: Nathaniel Whittemore Next up, moving over to markets, Oracle has shaken off negative sentiment with a strong earnings report. Coming into this week, the latest reports from Oracle was thousands of imminent layoffs to help fund their massive capex spend. A big part of the concern was that revenues would lag spending as data centers come online. Tuesday’s earnings call went a long way to settling those fears. Co-CEO Clay McGork reported that 400 megawatts of capacity had been delivered in the previous quarter, with 90% of that capacity delivered on time. Revenue related to server rental is up 84% year-over to reach $4.9 billion for the quarter. That growth rate was 16 percentage points higher than the previous quarter and beat analyst expectations by 5 points, demonstrating that demand is still accelerating. Oracle revenue grew 22% compared to last year, coming in at $17.2 billion. Oracle also noted that they wouldn’t need to raise more money to fulfill their obligations, noting most of the equipment needed is either funded up front via customer prepayments So Oracle can purchase the GPUs or the customer buys the GPUs and supplies them to Oracle. The stock gained 8% in after-hours trading, beginning to reverse the trend that saw the stock price cut in half since last September when the OpenAI deal was signed. Contrarian Curse on X writes, I thought Oracle did a good job on the call. They did paint a clean picture of why it’s not so easy to just slap AI everywhere. The only wrappers that are safe are ones that are embedded onto sticky platforms and workflows, and Oracle fits the bill. McGork spoke extensively on the call about why AI isn’t killing enterprise SaaS. One of the quotes, I’ve not yet met a customer who tells me they’re ready to give away their retail merchandising system, their core banking system, demand deposit accounting systems, Electronic health record systems, and that sub-small cobbling together of niche AI features are going to replace all of that overnight. Yes, we think AI is disruptive, but we think we’re the disruptor because we’re actually embedding the AI right into our applications at no additional charge. Overall, it seems like the market responded to the new co-CEO voice on the call. (Time 0:06:18)
  • Amazon Blocks Perplexity Shopping Agents Temporarily
    • A court granted Amazon a temporary injunction blocking Perplexity’s shopping agents for allegedly accessing password-protected accounts without Amazon’s authorization.
    • The ruling may set a precedent allowing marketplaces to force first-party shopping agents and restrict third-party agent competition. Transcript: Nathaniel Whittemore Amazon has won a court order blocking perplexity shopping agents from their platform. Last November, Amazon filed a lawsuit against perplexity, claiming their bots had fraudulently accessed the Amazon marketplace in breach of terms of service. The allegation was that perplexity was misrepresenting the nature of the traffic to circumvent web scraping controls. Amazon noted that Perplexity’s agents take control of a user’s account, arguing that this poses a serious security risk. Perplexity, meanwhile, argued that their bots were acting on behalf of users and should be treated identically to human traffic. On Tuesday, a judge granted a temporary injunction to prohibit the activity ahead of trial. They wrote in their decision, Amazon has provided strong evidence that perplexity through its Comet browser accesses with the Amazon user’s permission but without authorization By Amazon the user’s password-protected account. Articulating the legal standard to issue an injunction, the judge added that Amazon has shown a likelihood of success on the merits of its claim. Now, as this case continues, it could have pretty significant ramifications for agentic shopping. Primarily, Amazon is arguing that they should have control over how users access their platform, including the right to block third-party agents. However, they also discussed the advertising implications of agentic traffic. Amazon said that perplexity’s agents were served ads, which led to contractual issues with advertisers who only pay for human impressions. If Amazon is successful, they could set a precedent where marketplace websites have the ability to force customers to use first-party shopping agents, which some think would be stifling Competition in the still-nascent vertical. Perplexity for their part says that they will quote, continue to fight for the right of internet users to choose whatever AI they want. (Time 0:08:13)
  • Google’s Agents First Workspace CLI
    • Google shipped an agents-first Google Workspace CLI to make Drive, Gmail, Calendar, Sheets, Docs, and Drive programmatically accessible to agents.
    • Justin Ponelt designed the CLI so agents get deterministic JSON, self-describing schemas, and safety rails instead of GUIs or heavy MCP layers. Transcript: Nathaniel Whittemore Despite how powerful some of these new models are, and how cool the Genie 3 demo is, the release that I have seen get by far the most chatter is the Google Workspace CLI. And this, of course, speaks to just how important the coding use case is right now in driving the AI industry forward. For those of you unfamiliar, CLI stands for Command Line Interface. It’s basically a text-based way to talk to a program through your terminal. CLIs have been around forever and are the backbone of how developers interact with tools. If you want to use Stripe or AWS or almost any other developer tool, there’s a CLI for it. You type something like Stripe create payment in terminal and it just works. CLIs recently have become even more important as the better portion of agentic coding has been happening inside the terminal through harnesses like Cloud Code and Codex. You’re not clicking around in some GUI, you’re sitting in the command line talking to an AI that can execute commands. So if you are an agent builder and you want to integrate a new vendor, the path of least resistance is that the vendor has a CLI and your coding agent already being in the terminal can just Run the commands. No new protocol to learn, no new integration layer to build. Now Google, of course, has a lot of tools and spaces that agents might want access to. Drive, Gmail, Calendar, Sheets, Dot, etc. And up until recently, a lot of folks were defaulting to use something called GOG CLI that was built by Peter Steinberger, the same guy who built OpenClaw. It was a very big deal then, when last week Google dropped the official Google Workspace CLI. Mickey on Twitter points out the enthusiasm. Your OpenClaw, Claude Cowork, and perplexity computer agents just got a bit more useful. Kanika explained the value in simple terms. Agents can instantly read and summarize emails, draft and send replies, schedule meetings automatically, search drive for files, create sheets from raw data, generate docs and Reports, organize drive files, all from one agent workflow. Matt Silverlock noted the surprise of the old-is feel of this. He writes, 2026 is the year of the Chex Notes CLI? And Leon on X reframes it this way. They write, Google isn’t shipping a CLI for developers. They’re shipping an API for agents that happens to also work for humans. Google’s Justin Ponelt, who built the CLI, wrote a long blog post about it called You Need to Rewrite Your CLI for AI Agents. He writes, I built a CLI for Google Workspace, agents first, not build a CLI that noticed agents were using it. From day one, the design assumptions were shaped by the fact that AI agents would be the primary consumers of every command, every flag, and every byte of output. CLIs are increasingly the lowest friction interface for AI agents to reach external systems. Agents don’t need GUIs. They need deterministic, machine-readable output, self-described schemas they can introspect at runtime, and safety rails against their own hallucinations. He then goes on to write a whole bunch about the technicals behind this. Interestingly, a couple days later, he also wrote a piece about why for some there had been a shift away from MCPs and back towards CLIs. And before we actually read what he had to say, there’s some evidence that this is a broader phenomenon. Latent Space’s SWIX recently ran a poll. Let’s say you are an agent builder and want to integrate a promising new vendor you found. What would you be happiest to see in the docs? Not based on Twitter hype you personally for your situation right now. The options were API, MCP, or skills.md. Out of 769 people voting, MCP was actually in last place with just 9.1%. A traditional API was number one with 39%, followed by CLI with 31.2%, and a skills.md markdown file at 20.5%. Swix points out there was a time in 2025 when MCP would have been the clear number one on this list. In his blog post, The MCP Abstraction Tax, Justin sums up the issue this way. Every layer, data to API to MCP, introduces an abstraction tax. Humans need simplified abstractions to manage cognitive load. LLMs can navigate a complex CLI via help and call precise APIs in seconds. MCP and CLIs optimize for different things. Understanding what each one costs you is more useful than picking a winner. For complex enterprise APIs, the fidelity loss at each layer compounds in ways that matter. Basically, he says every protocol layer between an agent and an API is a tax on fidelity. That tax is sometimes worth paying, but you should understand what you’re giving up at each layer because the cost compounds. Kanika again sums it up this way. Most AI integrations use MCP servers, but MCP loads tons of tools into the context window. One developer measured 142 tools loaded, 37,000 tokens consumed, and 20% of context gone before work even starts. The CLI solves this differently. Instead of loading tools into context, the agent simply runs commands like gws drive files list. The CLI returns JSON and the agent continues. No context window tax. The takeaway is not that CLI is always better than MCP, but more that we’re still in the midst of the AI tooling transition. Everyone right now continues to experiment as things evolve with how to use old tools and systems, repurpose for agents, versus building new layers of infrastructure. (Time 0:14:55)
  • Provide CLIs Or Clean APIs For Agent Integrations
    • If you build agent integrations, provide a CLI or deterministic API outputs so agents can call commands directly and parse JSON responses.
    • Poll data showed developers prefer APIs (39%) and CLIs (31%) over MCPs (9%), so prioritize low-friction machine interfaces. Transcript: Nathaniel Whittemore No new protocol to learn, no new integration layer to build. Now Google, of course, has a lot of tools and spaces that agents might want access to. Drive, Gmail, Calendar, Sheets, Dot, etc. And up until recently, a lot of folks were defaulting to use something called GOG CLI that was built by Peter Steinberger, the same guy who built OpenClaw. It was a very big deal then, when last week Google dropped the official Google Workspace CLI. Mickey on Twitter points out the enthusiasm. Your OpenClaw, Claude Cowork, and perplexity computer agents just got a bit more useful. Kanika explained the value in simple terms. Agents can instantly read and summarize emails, draft and send replies, schedule meetings automatically, search drive for files, create sheets from raw data, generate docs and Reports, organize drive files, all from one agent workflow. Matt Silverlock noted the surprise of the old-is feel of this. He writes, 2026 is the year of the Chex Notes CLI? And Leon on X reframes it this way. They write, Google isn’t shipping a CLI for developers. They’re shipping an API for agents that happens to also work for humans. Google’s Justin Ponelt, who built the CLI, wrote a long blog post about it called You Need to Rewrite Your CLI for AI Agents. He writes, I built a CLI for Google Workspace, agents first, not build a CLI that noticed agents were using it. From day one, the design assumptions were shaped by the fact that AI agents would be the primary consumers of every command, every flag, and every byte of output. CLIs are increasingly the lowest friction interface for AI agents to reach external systems. Agents don’t need GUIs. They need deterministic, machine-readable output, self-described schemas they can introspect at runtime, and safety rails against their own hallucinations. He then goes on to write a whole bunch about the technicals behind this. Interestingly, a couple days later, he also wrote a piece about why for some there had been a shift away from MCPs and back towards CLIs. And before we actually read what he had to say, there’s some evidence that this is a broader phenomenon. Latent Space’s SWIX recently ran a poll. Let’s say you are an agent builder and want to integrate a promising new vendor you found. What would you be happiest to see in the docs? Not based on Twitter hype you personally for your situation right now. The options were API, MCP, or skills.md. Out of 769 people voting, MCP was actually in last place with just 9.1%. A traditional API was number one with 39%, followed by CLI with 31.2%, and a skills.md markdown file at 20.5%. Swix points out there was a time in 2025 when MCP would have been the clear number one on this list. In his blog post, The MCP Abstraction Tax, Justin sums up the issue this way. Every layer, data to API to MCP, introduces an abstraction tax. Humans need simplified abstractions to manage cognitive load. LLMs can navigate a complex CLI via help and call precise APIs in seconds. MCP and CLIs optimize for different things. Understanding what each one costs you is more useful than picking a winner. For complex enterprise APIs, the fidelity loss at each layer compounds in ways that matter. Basically, he says every protocol layer between an agent and an API is a tax on fidelity. That tax is sometimes worth paying, but you should understand what you’re giving up at each layer because the cost compounds. Kanika again sums it up this way. Most AI integrations use MCP servers, but MCP loads tons of tools into the context window. One developer measured 142 tools loaded, 37,000 tokens consumed, and 20% of context gone before work even starts. The CLI solves this differently. Instead of loading tools into context, the agent simply runs commands like gws drive files list. The CLI returns JSON and the agent continues. (Time 0:15:53)
  • Abstraction Tax Makes MCPs Costly For Agents
    • Every protocol layer between an agent and data (API, MCP, etc.) imposes an abstraction tax that reduces fidelity and increases context costs.
    • Justin Ponelt argues CLIs avoid context-window taxes by returning JSON so agents don’t preload huge tool contexts into LLM prompts. Transcript: Nathaniel Whittemore In his blog post, The MCP Abstraction Tax, Justin sums up the issue this way. Every layer, data to API to MCP, introduces an abstraction tax. Humans need simplified abstractions to manage cognitive load. LLMs can navigate a complex CLI via help and call precise APIs in seconds. MCP and CLIs optimize for different things. Understanding what each one costs you is more useful than picking a winner. For complex enterprise APIs, the fidelity loss at each layer compounds in ways that matter. Basically, he says every protocol layer between an agent and an API is a tax on fidelity. That tax is sometimes worth paying, but you should understand what you’re giving up at each layer because the cost compounds. Kanika again sums it up this way. Most AI integrations use MCP servers, but MCP loads tons of tools into the context window. One developer measured 142 tools loaded, 37,000 tokens consumed, and 20% of context gone before work even starts. The CLI solves this differently. Instead of loading tools into context, the agent simply runs commands like gws drive files list. The CLI returns JSON and the agent continues. No context window tax. The takeaway is not that CLI is always better than MCP, but more that we’re still in the midst of the AI tooling transition. (Time 0:18:24)
  • Gemini Leverages Workspace Context As A Strategic Moat
    • Google is using Gemini to tightly integrate Workspace context (documents, sheets, drive files) into generative features like Docs drafts and Sheet computations.
    • New Workspace features let users choose sources so Gemini-grounded outputs pull from a user’s Drive, improving relevance over generic models. Transcript: Nathaniel Whittemore Google AI Studios’ Logan Kilpatrick writes, Introducing the new Gemini-powered Docs, Sheets, Slides, and Drive experience featuring AI overviews, fully editable AI-made slides, And new grounding sources to make writing Docs context-aware. Sundar Pichai announced it this way, new Gemini updates to make Google Workspace more personal, helpful, and collaborative. Choose your sources and create a doc draft in seconds, build complex sheets nine times faster, or generate on-brand slide layouts with a simple prompt. Plus, Drive now generates summarized answers right at the top of your search results, so no more digging through folders. The blog post about this pitches it as a speed thing, but I Thank you. Actually think that there’s something else going on here. The post reads, dots and uncover useful insights while keeping your information safeguarded. When you look at the specific examples, though, a lot of the focus is on better access to the context that makes Google so powerful. So when you click on Create a Document with Gemini, you’re going to be able to select the sources in your Google ecosystem that it can pull from. And it’s that sort of integration that makes the experience so much smoother and hopefully makes the content on the other side that much better. The spreadsheet example they have asks for help tracking income for a particular month and again can pull from relevant sources like previous spreadsheets that live in Google Drive. Point being that while they’re pitching it as a speed play, the underlying idea here is better integrating the context that makes doing things from within your Google Workspace so Much more valuable. The sum totality of the documents that you have in your Google Workspace is something that Anthropic and OpenAI can’t compete with. It is a major advantage for Google and for Gemini. But only if they make that context accessible, and that I think is what this update is about. (Time 0:20:10)
  • Embedding 2 Removes Multimodal Search Friction
    • Gemini Embedding 2 is natively multimodal, letting search retrieve images, diagrams, screenshots, and text without converting them to captions first.
    • This improves retrieval for mixed-content knowledge bases, surfacing relevant slides, screenshots, or documents for queries like UI redesign discussions. Transcript: Nathaniel Whittemore Embeddings are basically the system that allows AI to find the right information. In traditional computing, search is done by keywords. If you search for buy a car, it’s going to look for those exact words. Embeddings, on the other hand, let the system understand that buy a car, purchase a vehicle, get a new ride are all basically the same request. Instead of matching words, they help AI match meaning. That means that when you’re building an AI system that has things like search or co-pilots looking through company documents or chatbots answering questions from knowledge bases, The system uses embeddings to quickly figure out which documents, files, or pieces of information are actually relevant. What makes embedding 2 a big update is that it is natively multimodal. So previously, if you had an image, a chart, or a slide, the system would have to convert it into text first, usually by generating a caption and then search using that. Multimodal embeddings remove that conversion step. Gemini embedding 2 can understand and retrieve images, diagrams, screenshots, text altogether. So if you asked a question, in a company knowledge base like, where did we talk about redesigning the checkout page? Theoretically, Embedding 2 could pull up a Slack conversation, a product spec document, a screenshot of the old UI, or a slide from a meeting, all as relevant sources. (Time 0:22:33)