AI Agents: Who Controls Agentic AI? | The AI Desk Ep. 37

This is The AI Desk, where today's signals reveal tomorrow's power. And today's signal is that AI has officially stopped asking, "How can I help?" And started asking... "Would you like me to just take over the entire workflow?" That is dramatic, Naya. It is also the news, Rowan. I contain multitudes. You contain caffeine. And insight. Occasionally in that order, because I was reading through the latest AI stories this week, and the vibe is not new chatbot. It is not even better assistant. It is, AI is becoming the operating layer. Search, video, documents, shopping, coding, research, finance, business workflows. Everywhere you look, AI is moving from answering to acting. That is the shift. And it feels fast, like uncomfortably fast, like when someone says, "Let's just do a quick rebrand," and suddenly, you are on hour six choosing between two identical shades of blue. That sounds personal. It was personal, and the blue was not even worth it. Noted. Mrs. Robinson's ghost is humming low. By the pool with a pink flamingo. They said... This episode is brought to you by MADCHITA and their new album WTF, Where The Forest? It's eco-pop engineered for the future, bold beats, global rhythms, and a message that actually matters. If you want music that hits your brain and your heart, explore WTF by MADCHITA, that's M-A-D-C-H-I-T-A. Streaming now on all major platforms. ... turn into plastic, it's fantastic. Floating in a sea of static people. Google just had I/O, and basically came out swinging with the whole agentic Gemini era. I was reading Google's own post, Google I/O 2026 News and Announcements, and then Wired's roundup, Everything Announced at Google I/O 2026. The pattern was impossible to miss. This was not one product launch. This was Google saying, "Gemini is going into search, shopping, building, productivity, video and agents." Google is not trying to make AI a feature. It is trying to make AI the interface. Exactly, and that word matters, interface. Because if AI becomes the interface, then you are not just opening apps anymore. You are asking for outcomes. Which means the user moves from command mode to delegation mode. That sounds like something a consultant would say right before charging me $12,000. It is still accurate. Annoyingly, because the Google announcements were everywhere. Gemini 3.5, Gemini Omni, AI Agents, Antigravity, AI inside search, AI inside creative tools, AI for shopping, AI for building. And Gemini Omni especially caught my attention because Google described it as starting with video and focusing on multi-modality, world understanding, editing and creation. That matters because video has always been one of the hardest creative formats. Yes, tell me about it, Rowan. Text was first, then images, then voice, now video. And once video becomes conversational, the entire creative economy shifts again. You are thinking about creators. Of course I am, because every creator I know is already exhausted. They are writing scripts, filming, editing, posting, clipping, captioning, resizing, analyzing performance, then doing it all again before the algorithm forgets they exist. A healthy lifestyle. Very peaceful. Very human. Very, "I cried into a ring light at 1:13 AM." That sounds oddly specific. No follow-up questions, please. But now, AI video tools are saying, "What if you could just describe the change?" Make this scene darker. Turn this into a Short. Change the background. Make the subject consistent. Add a product shot. Reframe this for vertical. Make me look like I slept. That last one may exceed current model capabilities. Rude. Accurate. Still rude. But the point is this is amazing, and also, it means the speed of content goes nuclear. Which brings us back to the larger theme. AI is not just making work easier. It is increasing the expected pace of work. Exactly. The reward for saving time is usually being asked to do more. That may be the most honest productivity sentence we have ever said. Thank you, Rowan. I suffered for it. But Google is only one part of the story. The other major story is OpenAI's research breakthrough. Yes, the math one. OpenAI announced that one of its models helped disprove a major conjecture connected to the planar unit distance problem, a problem first posed by Paul Erdős in 1946. Okay, pause, because when I first read Scientific American's article, OpenAI announces AI's biggest math breakthrough yet- I had two reactions. First reaction, incredible. Second reaction, I understood maybe 6% of the math. That is probably more than most of the internet. Thank you. I will be adding, understands 6% of advanced discrete geometry to my bio. Please do not. Too late. But seriously, the significance is clear. This is not AI summarizing a Wikipedia page. This is not AI writing a mediocre email that starts with, "I hope this finds you well." This is AI producing something new enough that mathematicians cared enough to verify, digest and discuss it. There is also an archive paper called, Remarks on the Disproof of the Unit Distance Conjecture, written by a group of mathematicians, including Noga Alon, Tim Gowers, Melanie Matchett Wood, and others. That paper gives a human verified version of the argument, and reflects on the result. And that is the part that made me sit up, because it is not just OpenAI says OpenAI did something. It is mathematicians saying, "Okay, this is real enough to engage with." That is the distinction. For years, people argued AI was just remixing. Now we are seeing more examples where models contribute to research, code, science and mathematics in ways that look less like autocomplete and more like discovery. And that changes the emotional tone, because if AI can help solve old math problems, then the conversation moves from, can it replace busy work to can it expand human knowledge? Yes, but it also raises a question. If AI becomes a research collaborator, who gets credit? There you are, Rowan. I knew you were going to make discovery legally complicated. It already is. Who owns the result? The company, the researchers, the model's developers, the humans who validate and refine it, the institution that funds the compute? So, even the miracle has paperwork. Every miracle has paperwork. That should be the title of your memoir. Noted. I would read it, by the way. You would skim it, Naya. Only the footnotes. That is where your personality lives. That is unexpectedly accurate. But this story matters, because it pushes AI deeper into elite knowledge work, math, science, engineering, drug discovery, materials research, the kind of work people assumed would stay human longer. Because we kept telling ourselves AI would do the boring stuff first. And it is, but it is also moving into the frontier stuff. So, the bottom gets compressed and the top gets accelerated. Exactly. That is the labor market story. Entry-level work gets automated, expert work gets amplified, and everyone in the middle has to redefine their value. That connects to our last episode on graduates, because if AI can handle beginner tasks and contribute to advanced research, the question becomes, where do humans train? That one keeps bothering me, because I remember being new at work and learning by doing the messy little tasks, the first draft, the bad spreadsheet, the client notes, the research memo where your manager writes, "Good direction, but clarify the point," which is corporate for, "I know what you meant, but the reader will not." That is how judgment gets built. Exactly. But if AI does all the beginner reps, humans may get fewer chances to become excellent. That is the unresolved problem. Okay, now let's talk about enterprise AI, because this is where things get sneaky. Anthropic and KPMG announced a global alliance to bring Claude into KPMG's core business. KPMG said its workforce of more than 276,000 people would get access, and that Claude would be embedded into KPMG digital gateway, starting with tax and legal use cases. That is huge, because when people think about AI adoption, they picture individuals using ChatGPT at home, someone asking for a meal plan, someone writing a breakup text, someone saying, "Make this email sound professional, but not dead inside." A common enterprise requirement. But the real adoption story is happening inside companies. Correct. AI is moving into audit, tax, legal, advisory, finance, spreadsheets, decks, emails, workflows, not as a novelty, as infrastructure. And that means AI is becoming less visible. Exactly. When AI is a separate chatbot, you notice it. When AI is inside Excel, Outlook, PowerPoint, Slack, your CRM, your project management system, and your company's internal database, you stop calling it AI. You just call it work. That is creepy, because it is true. The most powerful technologies disappear into routine. Like wifi or autocorrect or that one coworker who somehow controls the shared calendar, and therefore everyone's life. Exactly. I respect that coworker. I fear them. I may also need them. Enterprise AI is not about flashy demos. It is about embedding into the boring systems where power actually lives. Audit workflows, legal review, financial modeling, client presentations, internal reports, hiring screens, compliance. And once AI is embedded there, it starts shaping decisions. Not replacing the CEO dramatically in a glass office. No, more like quietly drafting the memo that the CEO reads. That is worse. It may be more realistic. Okay, so we have Google making AI the interface. Open AI showing AI can contribute to discovery, anthropic embedding into enterprise workflows. What is the fourth story? Regulations. Ah. The part where humanity realizes the plane is already in the air and starts debating seat belts. Exactly. The Associated Press reported that Pope Leo XIV called for robust regulation of AI and urged developers to prioritize the common good, rather than profit, focusing on AI's impact on work, war, and human dignity. That was one of the most interesting stories to me, because it shows AI is no longer just a tech topic. It is a moral topic, a labor topic, a war topic, a spiritual topic. AI has escaped the tech section. Yes, and that might be the real headline. AI is now in every section, business, politics, culture, religion, education, defense, dating, parenting, everything. Which means, the debate can no longer be left only to engineers and investors. Exactly. Because engineers will ask, "Can we build it?" Investors will ask, "Can it scale?" But society has to ask, "What does this do to people?" And whether we can still say no. Can we? Sometimes, but usually we say no after we have already become dependent. Dark. Historically supported. Annoying. Also historically supported. (laughs) You are very hard to flirt with when you're being this correct. You were flirting? Do not make me regret it. I would never. Smooth recovery. Thank you, Naya. But the regulations story is complicated, because nobody wants to be the country that slows down while everyone else accelerates. That is the arms race problem. And that phrase keeps showing up everywhere, AI arms race, compute race, talent race, agent race, safety race. Every story feels like a race. Because speed has become the strategy. But speed is not wisdom. No. Speed is often how systems outrun accountability. That is very AI-desk. Thank you. Okay, but I want to push back, because if we only frame this as danger, we miss why people are adopting it so fast. Go ahead. The reason AI is moving this quickly is because it is useful, painfully useful. Google's agents are useful. AI search is useful. AI video is useful. Claude in business tools is useful. Open AI helping with math is useful. That is why this is not like some random tech fad. People aren't being tricked into using AI. They are being pulled in because it solves real friction. I agree. That is why the debate cannot be AI good or AI bad. Right. The real debate is, who gets the leverage? Exactly. Does AI give leverage to individuals or does it concentrate leverage inside platforms? Does it help workers do better work or does it let companies hire fewer workers? Does it expand science or does it privatize discovery? Does it make creativity easier or does it flood every feed with synthetic noise? Does it make search smarter or does it make truth more dependent on whoever owns the assistant? Okay, now we're cooking. That also connects to AI search. Yes. I was reading Business Insider's piece on what marketers think about Google's big search and AI changes, and the anxiety was very clear. If AI agents can plan, recommend, shop, compare, book and buy, then brands are not just competing for clicks anymore. They are competing to be chosen by the assistant. That changes the funnel. It changes the whole room, because imagine you are shopping for running shoes. Old internet, you search, compare links, read reviews, open 15 tabs, get overwhelmed, maybe buy the pair with the least suspicious five-star reviews. And the most tolerable return policy. Exactly. New internet, you ask an AI assistant, "Find me running shoes for bad knees under $140 that work for city sidewalks and do not look tragic." A fair request. A necessary request. And the assistant gives you three options. Now the question is not, "Did my brand rank on page one?" The question is, "Did the assistant mention me at all?" That is a different power structure. And it's not theoretical. Researchers are already studying this. One archive paper titled, Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact, looked at AI overviews and found that these systems are changing how sources get selected, summarized, and monetized. There was also, How Generative AI Disrupts Search, another archive study comparing traditional Google search, AI overviews and Gemini results. Which gets to the uncomfortable part. AI search may feel convenient to users, but for publishers, creators, journalists, review sites, recipe blogs, forums, and niche experts, it can feel like someone read your work, summarized it, and then stood between you and your audience. That is the publisher anxiety. And honestly, I get it. I have searched for a recipe and never clicked through because the answer was right there. You cook? Occasionally. (laughs) Define occasionally. I own cumin. That is not cooking. It is a signal of intent. Fair. But this is the bigger point, every story this week is about AI crossing a threshold. From chatbot to agent, from content tool to creative system, from assistant to enterprise infrastructure, from search engine to decision layer, from research helper to discovery partner. And from tech product to social force. Yes, that is why this week feels different. It is not one announcement, it is the pattern. The pattern is that AI is moving closer to action. That's the episode title, AI Isn't Waiting For Permission Anymore. Exactly. Because the OLD AI waited for a prompt. The new AI is being designed to anticipate, act, coordinate, and complete. And that changes the risk model. Explain. A chatbot answer can be wrong. An agentic system can be wrong and do something. Oh, that is much worse. Yes. If a chatbot gives bad advice, you may ignore it. If an agent books the wrong trip, sends the wrong email, edits the wrong file, trades the wrong asset, schedules the wrong meeting, or summarizes the wrong source, the error has consequences. This is where I admit something. (laughs) Go on. I already get nervous when autofill changes one word in a text. As you should. One time I meant to write, "That sounds great," and autocorrect gave me, "That sounds greasy." To whom? A potential sponsor. How did they respond? They said, "Interesting feedback." That is devastating. Exactly. Now, imagine that energy, but the AI is booking your travel, emailing your boss, and editing a client deck. So the more useful AI gets, the more dangerous mistakes become. Correct. Capability increases value. It also increases blast radius. That is a brutal phrase. It is common in systems thinking. Of course it is. This is why verification becomes central. Not optional, central. But users hate verification. They do, because verification feels like homework, and the whole point of AI was to reduce homework. That is the tension. The better the assistant feels, the more tempting it is to trust it. And the more it acts, the more expensive that trust can become. Exactly. So how should people think about this news cycle? What is the practical takeaway? First, assume AI is moving from answer to action. If you are a worker, creator, student, founder, or manager, do not just learn how to ask AI questions, learn how to supervise AI workflows. That is good. Not prompting, supervising. Second, build verification into everything, sources, review steps, approval gates, human judgment. Especially if the AI is touching money, health, legal decisions, hiring, publishing, or anything public. Exactly. Third, pay attention to where your work lives. If your assistant has your memory, files, workflows, clients, and habits, switching costs increase. That connects to the platform dependency story. Yes. Fourth, understand that AI advantage is becoming contextual. The winner is not just the person who uses AI, it is the person who knows what good looks like. Taste, judgment, domain knowledge. Exactly. So, the least useful human in the AI era is the one who just accepts the output. Yes. The most useful human is the one who can direct it, question it, improve it, and know when to reject it. That is the skill. Okay, but I want to end on something human, because this week's AI news is huge and everyone is going to talk about models, agents, compute, regulation, enterprise deployments, search, and math breakthroughs. But underneath all of that is a very simple feeling. People are watching AI move into more parts of life and asking, "Am I being helped or am I being replaced?" Or managed. Yes, helped, replaced, or managed. That is the triangle. And the answer may be different depending on who you are. For a creator, AI video might feel like freedom. For a junior employee, AI agents might feel like the ladder disappearing. For a company, enterprise AI might feel like efficiency. For a regulator, it might feel like a system moving faster than the law. For a scientist, AI math might feel like a new collaborator. For a publisher, AI search might feel like the traffic disappearing. For a parent, AI homework help might feel like relief and panic at the same time. Same technology, different consequences. That is what makes this era so hard to talk about. There is no single AI story. There is a power map. (laughs) Of course you would say that. It is accurate. It is, and this week, the power map changed. Google wants AI to become the interface. OpenAI wants AI to become a discovery engine. Anthropic wants AI inside the enterprise nervous system. Regulators and moral leaders are trying to catch up before the systems become too embedded to challenge. And users are caught between convenience and control. Which is basically the entire internet, but faster and weirder. Yes. So the question is not, is AI coming? It is here. The question is, when AI starts acting for us, who is it really acting for? That is the question. And if your AI assistant starts making decisions, booking things, writing things, editing things, shopping for you, researching for you, and managing your work ... Mrs. Robinson's ghost is humming low. By the pool with a pink flamingo. They said... This episode is brought to you by MADCHITA and their new album, WTF, Where The Forest? It's eco-pop engineered for the future. Bold beats, global rhythms, and a message that actually matters. If you want music that hits your brain and your heart, explore WTF by MADCHITA. That's M-A-D-C-H-I-T-A. Streaming now on all major platforms. Plastics, it's fantastic. Floating in a sea of static. Maybe do not just ask if it is smart, ask, "Who gave it permission?" And, "Who benefits when you stop asking?" There it is. This is the AI Desk. Where today's signals reveal tomorrow's power.