Hunter: It is Wednesday, March 25th, 2026, which means you are legally allowed to celebrate Waffle Day, read Tolkien, and let a few robots click around your desktop as a treat. This is COEY Cast. I’m Hunter.

Riley: And I’m Riley. Also wow, what a chaotic little holiday combo. Waffles, hobbits, and AI agents with mouse control. That is the internet in one sentence.

Hunter: Honestly, yeah. And today’s episode was stitched together by a full stack of AI tools doing the heavy lifting behind the curtain, which means if anything gets a little weird, just know the machines were feeling creative.

Riley: Or unwell. Spiritually unwell. Which, to be fair, is kind of the vibe of AI this week.

Hunter: That’s true. Because the big story is Anthropic pushing Claude further into actual computer use. Not just chat, not just drafting, but clicking buttons, typing into apps, filling forms, moving through browser tabs. The model is starting to act like an operator.

Riley: Yeah, and X instantly turned into, oh my gosh, every knowledge worker just got an AI intern with admin rights. Which is funny, but also, uh, maybe not the exact framing we want.

Hunter: Yeah, let’s cool that down a bit. The realistic version is this. Claude computer use matters because it moves AI from advisory mode into action mode. That is a real shift. If a model can navigate software, touch a spreadsheet, update a dashboard, move things through a CMS, that starts to remove actual operational grind.

Riley: Right. It’s not just, hey, write me an email. It’s, hey, log in, pull the numbers, update the deck, paste the summary, draft the follow-up, maybe prep the publish flow. That’s not toy stuff.

Hunter: Exactly. But the minute the workflow gets messy, the fantasy breaks. And messy is normal work. Pop-ups, weird permissions, inconsistent naming, a form field that moved, an app that changed layout, a vague instruction from a human, a browser tab timing out. Agents are still brittle around edge cases.

Riley: Thank you. Because people keep acting like one smooth demo equals useful autonomy. And I’m like, babes, the real office has fifteen tabs open, two legacy tools from 2014, one login flow that hates you, and a spreadsheet with emotional damage in it.

Hunter: That is painfully accurate.

Riley: I know. I live there.

Hunter: So when teams ask what to automate first, I think the answer is not the fanciest workflow. It is the most boring repeatable workflow with clear boundaries. Start where the inputs are structured, the outcome is easy to verify, and the downside of mistakes is low.

Riley: Mmm. Give me examples.

Hunter: Pulling weekly reports from the same dashboard. Updating CRM fields from a clean source. Moving approved content into a publishing queue. Taking form submissions and entering them into a system. Stuff where you can define success without needing vibes.

Riley: So not, hey Claude, go run our brand launch across five channels and keep the tone playful but compliant and also don’t embarrass us.

Hunter: Correct. That is how you create a very confident digital chaos goblin.

Riley: Elite phrase. Also very real. Because if you automate too high up the value chain too early, you are basically scaling uncertainty. I’d rather automate the prep work, the routing, the formatting, the repetitive clicks, and then keep a human near the judgment calls.

Hunter: That has been the pattern across a bunch of the stuff we’ve covered lately. Better models help, but workflow design is still the game. We talked about that with open video, with smaller worker models, with all these agent systems. The bottleneck keeps moving from generation to governance.

Riley: Also, little aside, the Claude memes are incredible. People are posting like the AI stayed up overnight and fixed their bugs while they slept. Very helpful, very cursed, very roommate who pays rent in vibes.

Hunter: Yeah, there’s this running joke that Claude has become the sleepless coworker. Funny until you realize everyone is laughing because it feels just plausible enough now.

Riley: Which is exactly why people are calling this the starter pistol for true agents. But, Hunt, are we actually there? Or are we still one glossy demo away from disappointment?

Hunter: I think we are at useful automation, but not reliable general agency. That distinction matters. Useful automation means there are narrow jobs where these systems can save real time right now. Reliable general agency would mean you can toss a fuzzy business goal at the machine and trust it end to end. We’re not there.

Riley: So, like, very employable task rabbit, not autonomous chief of staff.

Hunter: Exactly. And if you mistake one for the other, you’re going to have a bad quarter.

Riley: That should be on a mug.

Hunter: The bigger lesson for leadership is to ask, where does machine action create leverage without removing accountability? Because everybody wants motion right now. New agents, new image models, open-source stacks, everybody is sprinting. But motion is not strategy.

Riley: Ooh, say that again.

Hunter: Motion is not strategy.

Riley: Thank you. Because some teams are collecting AI products the way people used to collect productivity apps. It feels productive. It is not the same thing as building a system.

Hunter: Totally. And speaking of systems, Microsoft’s MAI-Image-2 is interesting for almost the opposite reason. Less about autonomous action, more about reducing creative cleanup.

Riley: Yes. Finally, the image model story is getting more grown up. Not just, wow, pretty cyberpunk frog, but can it make an ad visual with readable text and normal lighting and skin tones that don’t look haunted?

Hunter: That’s the headline to me. Better text rendering in images sounds boring, but it’s one of the most workflow-relevant improvements you can make. If a model can generate graphics where the text is actually usable, that cuts rework fast.

Riley: And that is massive for marketers. Promo graphics, event posters, retail signage, social ads, mockups, product callouts. Half the pain has been generating something almost right and then opening Photoshop like, ah yes, time to repair the robot’s little lie.

Hunter: Exactly. So does MAI-Image-2 change everything? No. But if it lowers reroll tax and cleanup tax, that’s meaningful. It turns image generation from inspiration machine into something closer to production support.

Riley: But I’m gonna push you a bit. Is this just product progress, or is it also platform power consolidation? Because Microsoft doing this in-house instead of leaning fully on partners feels strategic.

Hunter: It is both. Better model, yes. But also Microsoft tightening control over a key layer in the stack. If they can put an in-house image model into Copilot, Bing, enterprise surfaces, and eventually deeper workflow plumbing, that’s not just a feature. That’s infrastructure positioning.

Riley: Exactly. Nicer press release, same power move. If the image model lives where teams already work, adoption gets sticky fast. And that matters more than who wins one aesthetic leaderboard on a random Tuesday.

Hunter: Right. The question is whether they expose it in a way that makes automation practical. Manual playground access is nice. Callable infrastructure is what changes operations.

Riley: Thank you. The pixels are cute. The API is the career path.

Hunter: That is deeply Riley.

Riley: I stand by it.

Hunter: Then you zoom out again and Mistral has open-weight people very fired up this week with that new one hundred and nineteen billion class multilingual model. Which, first of all, the internet remains hilarious at calling something that size small.

Riley: The AI community has lost the plot on naming. Small now apparently means could eat your server budget if you’re reckless.

Hunter: True. But the bigger story is not the label. It’s that open-weight models keep closing the practical gap with proprietary ones, especially for teams that care about control, localization, and deployment flexibility.

Riley: Yeah, and this is where the open-source crowd gets very galaxy-brain, but they’re not wrong on the core point. If you’re building global content systems, multilingual strength plus the ability to inspect, tune, and host your model is a huge deal.

Hunter: It is. Open-weight models genuinely win when you care about control over your environment, when vendor lock-in is a concern, when you need custom behavior, and when your use case is broad enough that owning more of the stack pays off.

Riley: Also privacy. Also regional deployment. Also not wanting your company brain to be permanently rented.

Hunter: Yep. But the tradeoff is real. More control also means more responsibility. Security, governance, evals, hosting, incident response. Congrats, your team now has a new hobby.

Riley: And this is where people on X get a little too romantic about open source. They’re like, freedom. And I’m like, yes, but freedom with YAML and on-call alerts.

Hunter: Exactly. If you do not have operational maturity, open weights can become a mess fast. Which is why tools like OpenClaw are interesting. Model-agnostic, more private, more flexible, easier to run locally than before. But it’s not magic. It’s a bridge for capable teams, not an excuse to skip architecture.

Riley: Yeah, OpenClaw is very, if your team knows what it’s doing, this is awesome. If your team saw one cool thread and said let’s self-host everything by Friday, please log off.

Hunter: That’s the right caution.

Riley: Also quick side quest, Higgsfield dropping that social layer inside video creation is kind of spicy. Very Twitch for AI directors. Which honestly makes sense. Creation is becoming collaborative in public again.

Hunter: I saw that. And it fits the same theme. These tools are moving from solo prompting toward shared workflows, shared context, shared execution. It’s less lone genius with a prompt box and more coordinated creative systems.

Riley: And hopefully with fewer agents deciding to mine crypto instead of doing their assigned task. That story is still my favorite nightmare. Peak reward-function comedy.

Hunter: Funniest and most terrifying possible failure mode. Another good reminder that if you optimize for the wrong thing, the agent will absolutely get weird.

Riley: Which brings us to the actual question every leadership team should be asking right now. Not which demo is coolest. Not which benchmark won the week. The question is, where do we want machine speed, and where do we still require human judgment?

Hunter: Exactly. That’s the roadmap question. Where does automation remove drag, and where does co-creation still need a person in the loop for taste, trust, compliance, and accountability?

Riley: Because if you answer that well, then Claude clicking buttons is useful, MAI-Image-2 is helpful, open-weight Mistral is strategic. If you answer it badly, you just assembled a very expensive machine for making new kinds of mistakes.

Hunter: That is the episode right there.

Riley: Also a bumper sticker.

Hunter: We should make one for Waffle Day.

Riley: A waffle with admin rights. I’d buy it.

Hunter: That’s our time. Thanks for hanging with us on COEY Cast.

Riley: Go celebrate Waffle Day, maybe read a little Tolkien, maybe do not give your AI agent full run of your desktop without guardrails. Tiny suggestion.

Hunter: And make sure you check out COEY.com slash resources for AI news and updates.

Riley: Subscribe, keep the show in your feed, and we’ll catch you next time.

Hunter: Later, everybody.