@thehypekevin

Meta, OpenAI and SpaceX AI each built an agent that wants your to-do list. Grok Bot lives on X, dots lives in ChatGPT, Muse lives in your WhatsApp. Which one gets the job? Tell me the craziest thing an agent has done for you. Narrated with an AI clone of my own voice. Music generated with Suno. #AIAgents #ChatGPT #Grok

Everybody has a list of things they don't want to do. The insurance claim, the package you still haven't returned, the dentist appointment you keep pushing back. Three of the biggest companies in AI want that list. Meta, Open AI and Elon's Spacex AI each built an agent that takes it off your hands while you get on with your life. Muse, Dots and Grokbot. And each one has a person out front sending it into the fight. So which one gets the job first? What they are, because all three are the same kind of thing. A chatbot answers you and then waits. These agents get their own computer in the cloud with a browser and files and they keep working after you close the app. They wake up on a schedule or when something new comes in and they ask before they spend your money or send anything in your name. Memory is the big one. Say an email lands about a concert. Your agent remembers you've wanted to see that band for a year, checks your calendar, finds the dinner it clashes with and asks if it should move the dinner. Mark Zuckerberg calls the goal personal super intelligence and all three are chasing it. They showed up within seven weeks. Grokbot in beta on August 11th, Muse on September 8th, dots on September 29th. Four days after that, the people behind two of them started trash talking. Teebo from open AI, pretty sure there are more dots than bots. Already Lauren from grokbot. I smell fear. Lauren Tan is the engineer you see posting grokbot all day and her own bot is a golden potato called potato, same as her handle. Grokbot is the agent for people who live on X since October 7th. It can search, read and monitor X posts, mentions what's trending right now and X is where the newest things get posted first. So ask it for the website designs everyone's sharing this week and a page in that style. It reads the posts and hands the build to a coding agent from cursor, which Spacex bought in August. The coding agent writes the page, checks a screenshot of its own work and opens a pull request for you to review. You can also run a whole team of bots. Lauren has another one, Sand Castle, that manages her releases and a team bot keeps one shared memory for everyone on your team. Then there's the brain. Grocbot ran on Groc 4.7, which scores 46 on the Artificial Analysis Intelligence Index. On October 7th, Elon said it would use whatever is most likely to give you the best outcome and named Claude Opus 5.5 Midjourney and Suno Opus 5.5 scores 58, top of the chart and it's already rolling out. Tibo's reply that night, goodbye groc models. Two days later, grocbot started getting its own email address. Ask your bot for an inbox and it can sign up for things and write to businesses as itself. It Comes with Kocher's 20 dollar pro plan or supergroc from 30. And on X, you can tag at bot under any post to give it a job. Lauren tested that on tebo. Yo at bot, send him a Hello World Python tutorial. So grocbot is the one that knows what the internet said this morning. Tebo's dots start from the opposite end. They know you. A dot starts with your chat GPT memory, every chat where you mentioned your kid, your job or your terrible taste in fonts. Ask it for an invite for your son's birthday on Saturday and it already knows he's turning 6 and wants dinosaurs. It hands the build to Codex on its cloud computer or on your own laptop. If you flip the switch that lets it in, Codex draws a Dinosaur and puts the invite up as a real website with an RSVP list that fills in as the parents answer. The dot asks if it can put the party on your Google Calendar. One tap. It also catches what you'd miss. Harry McCracken at Fast Company had a speaking event where the confirmation email and his calendar didn't match. His dot spotted it and told him. He says it was the first time AI saved his bacon. New on October 9th, you can set up a dot from your phone and start Codex jobs from there too. You can call your dot, hang up and it keeps working. For now, you get one dot And open AI says more are coming, which Grokbot already lets you do. Open AI has been shipping every day because on October 4th, Teebo promised 28 days of meaningful improvements to Codex and Chat GPT's work tools or reset everyone's usage limits. Each dot gets a name and a look. And Teebo's own dot is called Infinity. It's a little yellow reset button. Wearing glasses, Lauren's review of the 28 days. I think we might need more resets. Tibo, if you're watching, I could use one too. The brain is GPT6 Astra at 53. The catch is the price, $100 a month, at least through chat GPT Pro. And Pro isn't available in the EU, Switzerland or the UK. Europe, the world leader in AI regulation, gets it once it's been properly regulated, which leaves the one person who never joined the fight. Alexander Wang runs Meta's AI and his Muse is his puppy Euler. Not a picture of a puppy. He describes it as the super intelligent manifestation of his real puppy Euler. While Lauren and Tibo traded jabs, Wang was posting Muse fan films. Meta doesn't need to win an argument on X because mews can live in WhatsApp and 3 billion people already open WhatsApp every month. That changes who it's for. Your mom isn't downloading an agent app, but she can text one forward Muse the school emails and it pulls the dates onto the family calendar, flags the tryout deadline, and fills a cart with the Supplies on the list, then it can pay. In the US, Muse checks out with Stripe's link and every purchase stops in the chat with the exact total until you say yes. Muse never sees your card number. Meta has lined up Walmart, Best Buy, Sephora and Instacart for Muse, an announced voice and its own email address. Wang's pitch is that simple, add it to a thread or forward its stuff and it handles it. Under the hood, it gets the same setup as the other two, its own computer in the cloud with a browser, memory you can edit and a separate model called Sentinel that checks everything going out before it leaves. And Muse is free. There's a twenty dollar plan and a hundred dollar plan if you use it a lot, but you can start on nothing. Its brain, Muse Spark 1.3, scores 48 between Grock's own model and open AI's. And Wang said in September that meta is releasing the most capable model it has ever trained pretty soon. So its card has an empty slot waiting. Catch for now, US and Canada only put the three cards side by side. Grokbot meets you on X and in your code with the top brain on the chart, a team of bots and its own inbox. For $20, dots meets you inside Chat GPT where your memory and Codex already live. For 100, Muse meets you in the group. Chat pays when you say yes and costs Nothing. You can already talk to grokbot and dots out loud. And meta has shown Muse doing it on stage. And sooner or later, all three end up in the same place, business that has to deal with them. Agents are getting their own email addresses. Grokbot has one. Meta says Muse is next. And I'd bet phone numbers come after that. So say your agent writes to a company about your mortgage. It's polite, it's thorough, three paragraphs, every detail. It hits send and the one who opens it on the other side is another agent. Businesses are getting agents too. Sierra builds those for them and it says its agents have already started calling other agents. Built on Sierra, Sam Altman spotted this loop back in 2,023. Someone writes bullet points, has chat GPT, expand them into a polite email, and the person who gets it has chat GPT condense it back into bullet points. He called it very strange. Now capture both sides as agents writing each other paragraphs neither of them needed. So on October 6th, Meta and Sierra announced the personal agent protocol nicknamed Poppy. And on the 9th, they published the draft. Open AI and Cloudflare joined the design work with Shopify, Stripe and Walmart, already in a business, puts one small file on its site that says where agents come in. Your agent arrives and signs in as itself so the business knows an agent is talking. You sign in on the business's own Page and grant it read or write and it gets nothing more. Anything that matters like a purchase comes back as an exact offer with a price. You approve that exact version and it happens once. Even if the agent retries, change the price and the approval breaks. In the live demo, Muse went to rockets website, found rockets agent through Poppy and worked on a mortgage pre approval. Two agents, no paragraphs in our front door video. This spec was still coming. Now it's here and grocbot isn't on the partner list yet and they won't stay on your screen. On September 23rd, Meta showed the Muse Charm, a little device you carry in your pocket and just talk to. Muse is coming to Meta's glasses too so you can talk to it on the go all day without taking out your phone. On October 2nd, Meta also gave Muse a way into your house. The Home Link, a little USB C plug that lets Muse talk to your TV, your speakers and anything else on your Wi-Fi with a web address. So if your coffee machine has one, congratulations, Muse makes coffee now. In an interview with sources on September 1st, Sam Altman described the devices open AI is thinking about something that belongs on a table, something that belongs in your pocket, something that belongs on your body, whatever that means. Both want the part of your life that happens away from a screen. I hear it's a big part Unfortunately for me, I wouldn't know. So which one gets the job? For most people? Muse. It's free, it works in WhatsApp, and it's the one I set up for everyone I know who isn't online all day. If you're deep in AI, it's dots or Grokbot. And right now I'm leaning Grokbot, mostly because I'm on X all day and an agent that can read X is a big deal for me. But I'm still using Codex and my dots every day, so I'm watching this one closely. On October 8th, Lauren typed GG and the jokes stopped. The next day, Open AI shipped a dots update. Wang still has a new model and the charm coming. Elon still has midjourney and Suno to plug in. And apple has an event on the 13th. These cards are going to change and I'll update them every time they do. Who's winning so far? Vote under the video and tell me in the comments the craziest thing a personal agent has done for you.

You asked for a Pause button. Your agent also decided what happens to the bag of coffee that was already packed, and nobody asked you. Matt Pocock's fix starts before any code exists. Skills by Matt Pocock (x.com/mattpocockuk): https://github.com/mattpocock/skills Narrated with an AI clone of my own voice. Music generated with Suno. #AIAgents #ClaudeCode #DevTools

Say you run a coffee subscription app and you type, let users pause their subscription. 10 minutes later, there's a pause button. The tests pass and it works. Then a customer pauses on the 29th and still gets a bag of coffee on the 1st because it was already roasted and packed. Somebody had to decide what happens to that bag. Your agent decided. It also decided whether the Bill stops today or at the end of the month, how long a pause can last and what the reminder email says. You wrote one sentence, so it filled every gap with a guess and kept going. Agents are polite like that. Couldn't you write a longer prompt? You'd only write down the decisions you thought of and the packed bag wasn't one of them. Matt turns the conversation around before any code exists. The agent interviews you and his skill for it is called grill me, a relentless interview about your plan. Picture the feature as a tree of decisions in the fog. Pause sits at the top. Under it hang billing, the packed bag and how long a pause lasts. And each of those has more decisions hanging off it. The grilling skill maps that tree, finds every decision whose parent is already settled, called the frontier, and asks you all of them in one numbered round. Each question comes with the agent's recommended answer. So a lot of the time you type yes, yes, no, end of the Month your answers settle a layer, a new frontier lights up and the fog moves back. Facts are the agent's job. If the answer is somewhere in the code, it sends a helper agent to go and read it. Decisions are yours and it can't start on the code until the frontier is empty and you agree you're on the same page. On a small change, Matt gets about six questions and around 20 on a normal one. While filming a lesson for his course, it once asked him 46. His post about it. Dude, chill, you're scaring the noobs.

If you pay for Claude Max, there's up to $200 a month of API credit in your settings that you probably haven't claimed, and it can never bill you. Small coding jobs go to Haiku 5.5, the big one to a dynamic workflow. Credits by Anthropic (x.com/ClaudeDevs). Model data: Artificial Analysis. Narrated with an AI clone of my own voice. Music generated with Suno. #Claude #ClaudeCode #AIAgents

If you pay for Claude Max, Anthropic started putting up to $200 a month into your account this week and you probably haven't claimed it. It won't stretch your cloud code limits by a single minute. So what's it for? It's API credit, $100 on Max 5 x, 200 on Max 20 x. You claim it in settings under billing by linking a console organization. It pays for the API, managed agents and any tool that asks for a cloud API key. When it runs out, the requests stop and your plan is never charged. So it's the first API key you can hand an agent and still sleep that night the same week. Haiku 5.5 completed the set. Opus is the heavy thinker, sonnet the Everyday Coder, and Haiku the fast one. You hand small, clear coding jobs on artificial analyses index, Opus scores 58, Sonnet 56, and Haiku 43. At the same token price, Haiku outscored Open AI's GPT6 Luna. Sonnet outscored GPT 6.1 Soul, and Opus beats GPT6 Astra for less than half the price. Put score against what a task really costs and Haiku lands so far left that Theo called it wild for an anthropic model. So when you rename one function and 40 files still call the old name, Opus plans the change, 10 haikus fix the files and the credit barely notices. Big jobs break somewhere else. Anthropic hid 70 bugs in 116,000 lines of code and sent one Claud agent in three Times it came back with 14, then 15, then 27, and every time it said done. One agent carries the whole job in one context window, like one backpack. When the bag fills up, it squashes everything into a summary, and each summary loses a detail until one of them is what it was looking for. It approves its own finds because it wrote them. And around bug 15, the pile that's left starts to look finished. On Friday, managed agents got a different setup. The lead agent writes the plan down as a program. Phase 1 gives each piece of the code its own agent with an empty backpack. Phase 2 hands every bug they report to a fresh agent whose job is to prove it wrong. And since it didn't write the claim, it has nothing to protect. Anthropic calls it a dynamic workflow. One of its engineers calls them Big Claud and the little clods. Same 70 bugs. The workflow found 66 on every run. It's one field on your agent, and the credit pays for it. The week's most watched launch was Claud Motion. You describe an explainer like this one, and Claude animates it out of code instead of a video model so you can change one number and the rest stays put. It's team and enterprise only, so Max doesn't get it. And I'm still making these the hard way. So what's the $200 for? The 40 file rename goes to Haiku, the big Job goes to a team that checks its own work. So done means 66 instead of 15. Neither one can send you a Bill. And whatever you don't spend this month, anthropic keeps.

Your AI agent says it's done. Anthropic checked: one agent found 15 of 70 hidden bugs. A team of agents found 66. Here's what changed. Dynamic workflows by Anthropic (x.com/ClaudeDevs), Michael Cohen (x.com/_hi_mc), Rohan Seth (x.com/karbon0x) Narrated with an AI clone of my own voice. Music generated with Suno. #ClaudeCode #AIAgents #DevTools

You ask your agent to check a big code base for bugs. It comes back in a few minutes and says it's done. How many did it Miss? Anthropic tested exactly that. They hid 70 bugs in 116,000 lines of code and sent one clawed agent in. Three times it found 14, then fifteen, then on a good day, 27. Each time it planted its flag and called the job done. The agent carries the whole job in one context window. Every file it reads goes in the same backpack. And when the backpack fills up, it summarizes to make room. Each summary drops a detail, so by file 400, it's lost track of what the job was. When it checks its own finds, it likes them because it wrote them. And somewhere around bug 15, the pile of what's left looks finished enough to stop. So why not hand it a bigger backpack? Then it summarizes later and drifts anyway. Why not let it hire helpers? Claude Code can already do that, but every helper reports back into the same head, and the lead agent still decides each next step one turn at a time. As of today, Claude managed agents can do something else. The lead agent writes the plan down as a program and thropic calls it a dynamic workflow. Phase 1 splits the code into pieces and gives each piece its own agent with an empty backpack. Phase 2 takes every bug they report and hands It to a fresh agent whose only job is to prove it wrong. That agent didn't write the claim, so it has no reason to defend it. The program passes results between phases so the plan never gets summarized away. And the lead agent, which one of the engineers calls Big Clod, stays free to talk to you while the little clods work. Same test, same 70 bugs. The workflow found 66. 3 runs 66 every time. I pointed one at my own code. The publisher that post these videos, five agents hunted and came back with 10 bugs. Ten skeptics tried to knock them down and nine survived. One of them puts every follow up post on X in the same minute. The whole run took three and a half minutes and about 1.3 million tokens to turn it on. In the API, you set one field on your agent. A run can start up to 1,000 agents over its life and every one of them spends tokens. So you give the session a budget and the run pauses when it hits it. In code, you can try the same thing right now. Ask it to use a workflow. So how much did your agent miss? If your code looks like anthropics, test about 50 of the 70. Even on a good day, give every piece a clean head and someone whose job is to disagree and the flag goes up at 66. Anthropic made their launch video with a Workflow, by the way, I made this one with one agent. So if you find a bug, you know why I'm Kevin. When something new drops in AI or tech, I take it apart here and tell you if it's worth the hype. Subscribe so you catch the next one.

Stop fixing the same AI agent mistake twice. pstack's newest skill reads your history, finds the mistakes your agents keep making and turns each one into a rule the repo enforces. pstack by Lauren Tan (x.com/poteto): https://github.com/cursor/plugins/tree/main/pstack Narrated with an AI clone of my own voice. Music generated with Suno. #AIAgents #Cursor #DevTools

Her newest skill closes the loop. It's called correct. It reads your recent commits, reverts and review comments, and finds the mistakes your agents keep making. Twice counts as a pattern, and our hide the duplicate work around qualifies 40 times over. Then it climbs the ladder from the top, the cart shape that makes duplicates impossible and every copy of the work around deleted where architecture can't do it. It reaches for a type, then a lint rule whose error names the fix, then a test and docs. Last, every new check has to fail on a real mistake from your history. So the new lint rule runs against last month's workaround and goes red. And it keeps a table in your agent's instructions of which rule is enforced by what.

Your agent can build a whole feature while you make a coffee. So what are you there for? Matt Pocock's skills have more stars on GitHub than React. Here's his whole flow, start to finish. Skills by Matt Pocock (x.com/mattpocockuk): https://github.com/mattpocock/skills Narrated with an AI clone of my own voice. Music generated with Suno. #AIAgents #ClaudeCode #DevTools

Your agent can build a whole feature while you make a coffee. So what are you there for? Matt Pocock has an answer. He teaches developers for a living. He made total type script and his answer is a folder of agent skills with more stars on Github than react. On October 4th, he shipped version 1.3. Say you run a coffee subscription app and you type, let users pause their subscription. 10 minutes later, there's a pause button. The tests pass and it works. Then a customer pauses on the 29th and still gets a bag of coffee on the 1st because it was already roasted and packed. Somebody had to decide what happens to that bag. Your agent decided. It also decided whether the Bill stops today or at the end of the month, how long a pause can last and what the reminder email says. You wrote one sentence, so it filled every gap with a guess and kept going. Agents are polite like that. Couldn't you write a longer prompt? You'd only write down the decisions you thought of and the packed bag wasn't one of them. Matt turns the conversation around before any code exists. The agent interviews you and his skill for it is called grill me, a relentless interview about your plan. Picture the feature as a tree of decisions in the fog. Pause sits at the top. Under it hang billing, the packed bag and how long a pause lasts. And each of Those has more decisions hanging off it. The grilling skill maps that tree, finds every decision whose parent is already settled, called the frontier, and asks you all of them in one numbered round. Each question comes with the agent's recommended answer. So a lot of the time you type yes, yes, no. End of the month, your answers settle a layer, a new frontier lights up and the fog moves back. Facts are the agent's job. If the answer is somewhere in the code, it sends a helper agent to go and read it. Decisions are yours and it can't start on the code until the frontier is empty and you agree you're on the same page. On a small change, Matt gets about six questions and around 20 on a normal one. While filming a lesson for his course, it once asked him 46, his post about it. Dude, chill, you're scaring the noobs. Halfway through, you mentioned customers can skip a month. The agent stops you. Is skip the same thing as pause? This version is called Grill with docks and it keeps a glossary in your repo while it grills. Pause gets one line, no deliveries and no charges until the resume date. Skip gets another one. Delivery missed billing carries on. Under each the words to avoid like hold and freeze. An agent dropped into a project has to guess its jargon, so it uses 20 words where one would do Matt once read out A real reply from Opus 5 that ended. It is genuinely load bearing though, so that's how you know it matters. Then he asked the obvious question. What does that mean? So there's a skill for that moment too. Type wait, what? And the agent pitches its last message again in plain English using the words from your glossary. Some questions won't settle by talking, like what the pause screen should look like for those ungrillable ones. The agent hands off to a fresh session, builds a throwaway prototype with a few screens that look nothing alike and you click through them and pick. Then the answer comes back to the grilling. When the fog is gone, you stay in the same chat and type to spec. No more questions. This time the agent writes up everything you settled, the problem, the fix, and a long numbered list of user stories. As a customer who pauses after my bag is packed, I still get that bag and my pause starts after it. First though, it agrees the seams with you. The places where the tests will poke at the feature. And the ideal number of seams is one for us, that's the billing module. The spec goes up as an issue in your tracker labeled ready for agent. The spec is the destination. The tickets are the journey. Why tickets? Because an agent gets worse the more it's holding somewhere past 150,000 tokens. It stops reasoning as sharply. And the Part before that is the Smart Zone, a term Matt says he stole from Dex Hothi. So to tickets cuts the spec into slices that each fit one fresh session. Each slice runs through every layer database to button so it works on its own. Ticket 1, pausing stops the next charge. Ticket 2, a resume date. Ticket 3, the packed bag. Ticket 4, the emails 3 and 4 can't start until one is done. The rule of thumb, guess the tokens a job needs and divide by 150,000. So a million token refractor becomes 7 tickets. Then the build type implement on ticket 1, clear the chat, implement ticket 2 so every ticket starts in a clean head. Or use implement spec new in 1.3, which builds the whole spec in one go. One agent per ready ticket, each in its own copy of the repo, all landing on one branch. Either way it works. Test first, a failing test at the agreed seam, then only enough code to pass it one slice at a time. The skill used to have a refactoring step too. Matt dropped it because the agent never ever, ever did it. Refactoring moved to review and review happens at a different desk. An agent that has written some code thinks it's fine because it wrote it. So Code Review hands the diff to two fresh agents. One reads it against the spec looking for anything missing, anything nobody asked for, and anything that looks Done but is wrong. The other reads it against your coding standards and a dozen classic code smells from Martin Fowler's refactoring. The agent that wrote the code never saw those standards. Models are bad at following a list of don'ts while they build. And a reviewer hunting for exactly those problems has nothing else on its mind. On ticket 3, the spec reviewer quotes our user story back. Pausing today cancels the packed bag when the fix goes up as a pull request. A skill called PR writes the description. It's adapted from Dex Horthe's show me skill, the smallest picture that explains the change proof from before and after. And a section called Merge Danger. Is this a one way door or a two way door? And how big is the blast radius? Ours adds a column to the subscriptions table. Two way door blast radius. Billing. The last step is the other big one in 1.3. Before you clear the chat, type retro. The agent reads back through the session it just had and suggests changes to its surroundings. The repo, the checks the instructions. That matters because your agent won't complain. It bumps into problems, keeps pushing until the feature works and the struggle stays buried in the transcript. Here it spent 10 minutes hunting for the billing code. So it suggests one line in your agent instructions that points at it. It called the payment provider directly twice, which your codebase never does so that Becomes a lint rule. A mistake a machine can spot gets a check, full stop. And the written coding standards keep only the judgment calls found an error, do a retro and make it impossible next time. People keep asking Matt to put retro on autopilot and he says don't because an agent fixing its own findings ends up chasing false positives. It only suggests and you pick which fixes to keep. A coffee app is a toy. So take something real. On September 25th, in the app Matt uses to make his courses, a spec went up with 41 user stories. Within four minutes, there were 10 tickets linked by what blocks. What? One ticket started the chain, five waited on it and the last one stayed open on purpose as research. All of it landed on one branch. 28 commits and almost 14,000 lines merged a little over two hours after the spec was written. Five days later, the next features pull request had its own merge. Danger section 2 way door blast radius deploy ordering. Isn't this waterfall with extra steps. That's the most common pushback. And Uncle Bob made the case on Matt's own live stream. Watch what happens to the spec though. Once the work ships, Matt closes the issue and the spec is gone. He calls a spec a projection of the decisions made during grilling. And by then, those decisions live in the code, the tests and the glossary. The prototypes keep it from Turning into waterfall and anything small enough for one session gets planned in one session. The other pushback says smart models don't need this much structure. One builder switched to P stack, the stack from our last video, because Matt's skills felt too defensive for today's models. Most of them cost nothing until you call them, though. Sixteen of the 27 skills in the plugin only exist when you type their name, so the agent never sees them. And in his tutorial, the whole installed set took up 660 tokens. He also deletes skills once the models catch up. Version 1.3 removed the merge conflict skill because the harness can handle it. So what's inside the repo with more stars than react? Open grill me. Which skills dot s H counts at over 1 million installs. It's one line long. Call the skill tool with grilling. The grilling skill it calls is 28 lines. And Matt calls the original grill me the most influential four sentences I've ever written. There's almost nothing in there to copy. The value is in the questions and only you can answer them. So what are you there for? The decisions. The agent can write every line of the pause feature and every call inside it still belongs to you. The Bill, the bag, the words skip. Matt thinks of agents as brilliant tactical programmers who need a strategic one above them. So the agent is the sergeant and you're the general. Our customer pauses on The 29th gets the bag that was already packed and pays nothing more until the resume date. You decided that in round two on your coffee break. I'm Kevin. When something new drops in AI or tech, I take it apart here and tell you if it's worth the hype. Subscribe so you catch the next one.

You can stop writing scrapers for your AI agents. Firecrawl's new Universal Scrape reads a website, a PDF, a LinkedIn profile and a podcast with one call. Universal Scrape by Firecrawl (x.com/ericciarla) Narrated with an AI clone of my own voice. Music generated with Suno. #AIAgents #DevTools #WebScraping

Your agent can write the code and book your flight. So why does reading up on one person still take it? Four different tools. You have a call tomorrow with a founder named Priya and your agent preps you. Her website needs a scraper. Her pitch deck is a PDF, so that's a parser. Her LinkedIn blocks scrapers, so that's an enrichment service with its own account. And she did an hour long podcast. So that's downloading audio and paying for a transcript. My own transcription script opens with a list of everything that broke on day 1, four API keys, four bills, and whichever one breaks will pick the morning of the call. Why not hand the agent a browser? LinkedIn shows it a login wall and a browser can't skim an hour of audio. Firecrawl shipped an answer today called Universal Scrape. Its scrape call already read websites and Pdfs. Now it takes the other two on the same API key. The LinkedIn link is the sneaky one because Firecrawl won't scrape LinkedIn. It passes the link to data companies you picked like Apollo in your order. The first one with Priya's record answers and the ones that come up empty charge nothing. Your agent never sees the login wall. The podcast has no link at all. So your agent searches Firecrawl's provider library called Alexandria for free and finds particle, which has podcast transcripts. The tool lists what to send, what comes back, and the price. So The agent calls the same scrape with the provider, the tool, and Priya's name. Back comes the transcript with the bit where she says the worst meetings are the ones that run long. Alexandria has over 200 providers and Firecrawl pays data providers for access, Wikipedia included. Next, they want to pay people for what they know. Update the Firecrawl CLI or MCP your agent uses, and it can search the whole library. So how many tools does it take to read up on Priya? One call and one key and the other three go back on the shelf. You walk into that call knowing to keep it short because she said so on a podcast, which is also why this video is under 2 minutes. I'm Kevin. When something new drops in AI or tech, I take it apart here and tell you if it's worth the hype. Subscribe so you catch the next one.

Before the merge: a judge, three models and a squad of verifier agents. After the merge: Lauren, tasting a sample of the commits in the morning. pstack by Lauren Tan (x.com/poteto): https://github.com/cursor/plugins/tree/main/pstack Narrated with an AI clone of my own voice. Music generated with Suno. #AIAgents #Cursor #DevTools

Then three models try to break it. The interrogate skill hands the diff to Opus, Soul, and groc separately, and the lead sorts what they find into act on, consider, noted, and dismissed. On full autopilot, Pstack sends a squad of verifier agents at every pull request, they click around the running app like real users, hunt for regressions, fix what they find, and go again until the PR can land. Wait. So who reviews all this before the merge? A judge, three models and a squad of verifiers. After the merge, Lauren. She reads the commit history in the morning and tastes a sample the way a chef can't taste every plate if two agents took the same shortcut. She adds a wire. The first night she let the agents merge on their own, she was scared she'd break production. Now she's sleeping so much better.

Your next customer is an AI agent. Here's what your app needs so it can sign in, get permission and pay, and how your own agents can use the same doors. AgentID by AgentMail (x.com/haakamaujla), PACT by Decagon (x.com/gramliu) Narrated with an AI clone of my own voice. Music generated with Suno. #AIAgents #DevTools #Startups

What does your app do when the next customer to sign up isn't a person? This week, four teams put out answers and if you build anything people pay for, all four are headed for your login page. You vibe coded a booking site for a yoga studio and Maya asks her AI agent to book her the Tuesday class. The agent opens your site and hits the login page. Block it like a bot and you just turned away a paying customer. Or let it type in Maya's password. And now a program can do anything she can, including changing her card. And your logs can't tell the two of them apart. First question, who is this? Agent Mail, which gives agents their own email inboxes just launched Agent ID sign in with Google for agents. It runs on Open ID Connect, the same standard as the Google button. So most off libraries already speak it. The agent signs in with its inbox and your app gets back a sign token that says actor type agent. If Maya allows it, the token carries her name and email too. Now you know it's a bot and whose bot it is. Second question, what did Maya agree to? Decagon just open sourced a protocol for it called pact. The agent shows its signed ID and sends Maya a link to your own login page where she ticks the boxes. Book classes, yes. Touch my card, no. The agent gets a token With only those permissions. And every action comes back with a signed receipt. So when the agent decides Maya would love the unlimited plan, your app already has the answer packed, sits on top of oauth, and the code is on Github under Apache 2. Third question, how does an agent find your booking page at all? Meta and Sierra are writing a standard for that, the Personal Agent protocol. With Shopify, Stripe, and Walmart on board. The agent reads what your business offers, browses as a guest, and asks the customer to sign in only when it needs her account. The first spec lands later this month, and with meta behind it, that's a lot of assistance that could come knocking. Last question, how does it pay without Maya's card? You've seen this one in our Cloudflare video. Your server answers with htttp 4:02, payment required and a price. The agent signs a payment and asks again, and the class is booked. Status code 4:00 2 was reserved for future use in 1997. The future took 29 years and it turned out to be a bot buying a yoga class. Now flip it around. If you run your own agents, all of this works for you too. Give your agent an agent mail inbox and it can sign itself up to apps that take Agent ID like firecrawl and torso without borrowing your Google account. And with 4:02, it pays for the one API call it needs instead of You buying a monthly plan? You'll forget to cancel. So what does your app do when Maya's agent shows up? It sees an agent and knows it's Maya's. It lets it book Tuesday, keeps it away from her card and gets paid. And nobody handed a bot their password today. That's four projects from four teams, and none of them plug into each other yet. Whoever wires them together first is going to have a very busy login page. I'm Kevin. When something new drops in AI or tech, I take it apart here and tell you if it's worth the hype. Subscribe so you catch the next one.

Your AI agent is hiding bugs, and it calls them temporary. A month later that comment lives in 40 files. Lauren Tan's fix is to stop snipping and start wiring the tree. pstack by Lauren Tan (x.com/poteto): https://github.com/cursor/plugins/tree/main/pstack Narrated with an AI clone of my own voice. Music generated with Suno. #AIAgents #Cursor #DevTools

Your shop has a bug. Sometimes the same hoodie shows up twice in the cart. You ask your agent to fix it. It can't work out why, so it adds a line that hides duplicates on the cart page with a comment, temporary workaround, safe to ignore. The tests pass, the cart looks right, and the bug is still there. Next week, another agent finds a doubled item in the wish list, sees that pattern next door, and copies it, comment and all. You catch it, explain it, and get another. You're absolutely right. A month later, the temporary work around lives in 40 files. Nothing lasts longer than a temporary fix. Lauren has a picture for this. Commits pouring into a codebase grow like a bonsai tree fast and in whatever direction they please. Every correction snips one branch. Tomorrow, the same wrong branch grows back somewhere else because agents copy whatever is already there. And a workaround like ours can spread through a whole code base in a couple of weeks. Bonsai growers wire the branches instead, so the tree grows into the shape by itself. Lauren calls it gardening, and a lot of those 25 PRs are exactly that.

Stop correcting your AI agent. Lauren Tan shipped about 2,500 PRs in one month and says each correction only once. Here's her whole system. pstack by Lauren Tan (x.com/poteto): https://github.com/cursor/plugins/tree/main/pstack Narrated with an AI clone of my own voice. Music generated with Suno. #AIAgents #Cursor #DevTools

How many times this week did you tell your AI agent the same thing? It says you're absolutely right, fixes it, and two days later, a different chat makes the exact same mistake. Lauren Tan stopped having that conversation in August. She shipped about 25 poll requests, around 80 a day, and her agents merged their own code. Lauren builds grokbot at Spacex's AI team and works on the react compiler. And Matt Pocock put her in his top 3 skill makers. Her whole system is public. It's a set of agent skills called pstack. And her newest skill is aimed at that exact conversation. Your shop has a bug. Sometimes the same hoodie shows up twice in the cart. You ask your agent to fix it. It can't work out why, so it adds a line that hides duplicates on the cart page with a comment, temporary workaround, safe to ignore. The tests pass, the cart looks right, and the bug is still there. Next week, another agent finds a doubled item in the wish list, sees that pattern next door and copies it, comment and all. You catch it, explain it, and get another, you're absolutely right. A month later, the temporary work around lives in 40 files. Nothing lasts longer than a temporary fix. Lauren has a picture for this. Commits pouring into a codebase grow like a bonsai tree fast and in whatever direction they please. Every correction snips one branch. Tomorrow, the same wrong branch. Grows back somewhere else because agents copy whatever is already there and a workaround like ours can spread through a whole codebase in a couple of weeks. Bonsai growers wire the branches instead so the tree grows into the shape by itself. Lauren calls it gardening and a lot of those 25 PRs are exactly that and the wires come in a ranking. The best one is architecture, design the code so the mistake can't happen at all. Next comes a lint rule or a test so CI catches it third, a skill or a rule the agent reads and dead last, a human reviewing the code to catch it, which she labels N G m. I not gonna make it for our cart. The top rung is fixing the shape of the data one line per product with a quantity so the cart can't hold the same hoodie twice and then deleting the work around everywhere. The second rung is a lint rule that flags code hiding duplicates and points at the real fix. Pstack's Fix root causes rule which tells the agent to trace every symptom back to its cause is rung 3 and you spotting the 40th copy and code review are rung 4. Why isn't the rule enough? Because agents take shortcuts, even the best ones. Pstack is written for an agent that sees only the files it opened, copies the nearest example and takes the shortest path that compiles that agent reads the work around And its comment. And the comment tells it the shortcut is fine. So Lauren makes the easy thing the right thing. The framework behind Grokbot has one way to do anything. Every feature gets its own folder. Lint rules block the wrong imports and code comments are banned because agents kept using them to excuse their workarounds. It drives humans a little crazy. The agents love it. Pstack is how she climbs that ladder. On every task, it's a free plugin for cursor with 50 skills and you only type one potato mode plus your task. It picks one of 23 playbooks like bug fix, feature performance, or shipping. And there's one for going to bed and waking up to everything merged. It copies that playbook, steps into a to do list and calls the other skills when a step needs them. About half the skills are single rules like delete before you add or after two fixes fail. Question the Assumption they share. The whole loop goes like this. Work comes in a bug report, a feature, a slow page. Potato mode picks the playbook. The agent reproduces or maps the problem, designs a fix, builds it, proves it in the running app, lets other models attack it and ships it. Anything that went wrong twice gets wired into the repo. So the next lap starts from a better tree. And Lauren doesn't type most of that work in. Grokbot watches her Slack X and bug tracker and turns New reports into projects that run this loop. She calls that the outer loop and it's how one person gets to 25 PRs in a month. Let's run our cart bug through it. Potato mode items show up twice in the cart. Reproduce it first. That's the bug fix playbook. And step one says the agent reproduces the bug itself in the running app. That used to be Lauren's job before verification skills. She was the meat proxy between her agent and Chrome's developer tools, carrying traces back and forth by hand. Pstack writes your project a verification skill instead with how to launch the app, check it's healthy, click through it and capture proof. Plus a map of where every feature lives. So the agent opens the shop, double clicks add to cart and gets two hoodies. Step 2 is finding the cause and guessing isn't allowed. The house skill sends read only agents through the code to trace the click from the button to the cart. The wise skill digs through old commits and chats and turns up last month's temporary workaround. The agent tests one theory at a time with real evidence until one survives a double click, sends the request twice and the cart happily stores both. The playbook has a line for workarounds too. A fix that might help is a hypothesis and it doesn't ship. The fix touches the button and the cart. So the architect skill steps in. It writes the code that Will use the fix first, then the types, and it never trusts a single design. Two agents design the fix in separate copies of the repo. One blocks the double click on the button. The other changes the cart so it can't hold the same product twice. A judge from a different AI lab scores them and the lead agent takes the cart design and keeps the button idea too. Grock writes the code and Opus 5.5 makes the hard calls. A failing test lands in the history. First two ads and one hoodie expected, then the fix on top. And the agent goes back to the same running app, double clicks again, and sees one hoodie. It compiles doesn't count as proof, and neither does a green unit test on its own. Then three models try to break it. The interrogate skill hands the diff to Opus Soul and groc separately, and the lead sorts what they find into act on consider, noted and dismissed. On full autopilot, Pstack sends a squad of verifier agents at every pull request, they click around the running app like real users, hunt for regressions, fix what they find, and go again until the PR can land. Wait. So who reviews all this before the merge? A judge, three models, and a squad of verifiers. After the merge, Lauren. She reads the commit history in the morning and tastes a sample the way a chef can't taste every plate if two agents took The same shortcut, she adds a wire. The first night she let the agents merge on their own, she was scared she'd break production. Now she's sleeping so much better. Her newest skill closes the loop. It's called correct. It reads your recent commits, reverts and review comments, and finds the mistakes your agents keep making twice counts as a pattern. And our hide the duplicate work around qualifies 40 times over. Then it climbs the ladder from the top, the cart shape that makes duplicates impossible and every copy of the work around deleted where architecture can't do it. It reaches for a type, then a lint rule whose error names the fix, then a test and docs. Last, every new check has to fail on a real mistake from your history. the new lint rule runs against last month's work around and goes red. And it keeps a table in your agent's instructions of which rule is enforced by what the same day. The architect skill Learned to spot traps like these before they get built. Two ways to do the same task lists you keep in sync by hand. Internals anyone can import. In a blind test against the old version, its design scored about a point higher out of eight. The stack also keeps getting shorter, which I did not expect. In September, Lauren cut 34 instructions that Opus 5.5 follows without being told. And she AB tested the second batch. Her review skill caught 142 of 154 planted problems without those lines and 139 with them. The rule behind it fits on a sticky note. If a lesson can be a lint rule, a check or a script, encode it and delete the instruction. Agents learn from the repo they're standing in. So that's where the lesson has to live. So how many times will you tell your agent about that? Work around once. Lauren's agents still make mistakes, but every mistake she sees twice turns into a wire. A cart that can't hold duplicates, a lint error that points at the real fix. The next agent that hits a weird bug finds nothing to copy. The tree still grows fast and now it grows in her shape. I'm Kevin. When something new drops in AI or tech, I take it apart here and tell you if it's worth the hype. Subscribe so you catch the next one.

Cloudflare wants to replace your entire stack. After Birthday Week, one small app might need one company, and your agent runs it from the terminal. Narrated with an AI clone of my own voice. Music generated with Suno. #Cloudflare #AIAgents #DevTools

Run your AI tests on any model, even a local one. e2e works with OpenRouter, Vercel's AI Gateway, a provider's API, a local model, or the ChatGPT, Copilot or SuperGrok plan you already pay for. Laptop, CI, real browsers, iOS and Android simulators. Narration uses an AI clone of my own voice. Music generated with Suno. #AIAgents #SoftwareTesting #DevTools

Your agents ship PRs all day and you're stuck checking every one. TesterArmy just open-sourced e2e so an agent tests them for you. Narration uses an AI clone of my own voice. Music generated with Suno. #AIAgents #SoftwareTesting #DevTools

TesterArmy's e2e saves the clicks an agent made once a check passes, then replays them on the next run without calling the model. Rename the button, and the agent steps back in to find it. Narration uses an AI clone of my own voice. Music generated with Suno. #AIAgents #SoftwareTesting #DevTools

Stop testing your AI agent's code by hand. TesterArmy's open-source e2e gives your agents a test loop and remembers the clicks that worked. Narrated with an AI clone of my own voice. Music generated with Suno. #AIAgents #SoftwareTesting #DevTools