Andrew Wilkinson interviews Dan Shipper, co-founder and CEO of Every, about agent-native architectures and how to build apps in an era where code is increasingly written by AI agents. # Channel Presentation Observations — "Agent-Native Apps: How to Build Apps After the End of Code" ## Format - **Conversation / pair-demo format**: two participants with a ">>" speaker marker convention throughout, indicating an interviewer–guest or host–guest dialogue rather than a solo tutorial. - **Live screen-sharing walkthrough**: opens with "Let me show my screen. Okay. Can you see this?" — demo-first structure. ## Intended Audience - **Hands-on "vibe coding" builders and AI-native entrepreneurs**: the discussion assumes familiarity with Claude Code, ChatGPT, Loom, Notion agents, Playwright, and APIs (Tavily, Ampify, Firecrawl) without explanation. - **Operators/executives interested in AI-driven company workflows**: extended segment on 2026 company planning via a Notion agent, meeting recording culture, and change management ("we've got 30 CEOs"). ## Claimed Scope (from title) - Title frames the channel topic as **building software in an "agent-native," post-code era** — the "end of code" framing positions AI-assisted development as the central theme. ## Demonstrated Presentation Patterns - **Practitioner credibility through live demos, including failures**: the guest shows a "Deal Hunter" analysis app that mislabels a known company ("highly speculative acquisition opportunity") and admits "I'm still dealing with hallucinations and data being inaccurate." - **Candid trade-off discussion**: "you can really get 95% of the way there in like 2 hours and then the final 5%... it goes on forever." - **Named ecosystem advocacy**: specific tools recommended — "Compound Engineering plugin," "agent native review," happy.engineering, "Claude Island," "/model opus plan," Sonnet 1M. - **Self-referential dogfooding**: "we released this guide today and... it has a button that says read with Claude." ## Self-Description Evidence - One speaker describes the business explicitly: > "we're the only subscription you need to stay at the edge of AI" and "our whole business is to do this because it filters for people who are just all in on trying to do it well." - Company scale stated: "we're now 20-ish people or whatever." - Companion speaker context: "we've got 30 CEOs" and references to "Tiny," implying a portfolio-company executive as the other participant. ## Notes on Distinguishing Claims vs. Demonstration - The "agent native" framework is a **creator claim** promoted via the guide; the video demonstrates enthusiasm and partial success, but the live demos shown do **not** fully validate the framework (both apps exhibit bugs and missing data during the demo). - No channel-wide mission statement beyond the title; audience and scope are inferred from demonstrated content and the quoted self-descriptions above. # Delivery Observation: "Agent-Native Apps: How to Build Apps After the End of Code" ## Format This is a **two-person conversational screen-share demo**, not a scripted presentation. Delivery is improvised, with live software demos, interruptions, and laughter. ## Tone - **Casual and enthusiastic**: "It's just like it's just going until it works," "Dude, I just got to say you're so lucky" - **Candid about flaws**: "It might be broken, but um but anyway, it's been not helpful yet"; "I'm still dealing with hallucinations and data being inaccurate" - **Superlative-heavy**: repeated intensifiers — "insane" (at [125], [673] area: "it makes the idea… totally unheard of and insane"), "it's so cool. It's so useful," "I'm just astounded" ## Recurring Verbal Habits - **"So, basically" as a transition phrase** — appears repeatedly: "So, basically, I built this as like an initial screen…"; "So basically the once you fill out…"; "So, so basically the once you have all your results…" - **"Like" as a filler**, dense throughout: "like an initial screen," "like they just spend an hour," "it's like a really good workflow" - **"Um" fillers** frequent in the demo sections, less so when discussing architecture - **"And so what…" / "And then…" chaining** to narrate demos step by step: "And then um it runs evaluation, generates a report, and then we can develop a deal memo based on it" ## Teaching/Demonstration Pattern - **Live walk-through with self-narration**: describes what the software is doing in real time ("It's going to go out and get um company background, any financials it can find") - **Pre-emptive expectation management**: repeatedly flags the tool's failures while demoing ("it's not accurate but what's shocking is just how far I was able to get") - **Concrete anecdote-driven explanations**: rather than abstract principles, the speakers illustrate points with stories ("And then it told me about something that had happened in that meeting that I would have had an opinion on") ## Recurring Themes in Vocabulary - **"Vibe coding"** used as established jargon: "the most infuriating thing about vibe coding" - **The "95% / last 5%" framing**, stated twice: "you can really get 95% of the way there in like 2 hours and then the final 5%. It just it goes on forever" and later "the last 5% has me been it's been me going in circles" - **Tool names dropped fluidly** without explanation: Claude, Tavali, Ampify, Firecrawl, Granola, Fireflies, Otter, Notion agents, Loom - **Colorful metaphors for change management**: "It's like nailing jello to the wall… When you've got a cruise ship going a certain direction, it's very difficult to turn it"; "I'm riding a bike and I was walking before… I have a helicopter" ## Differentiating Speakers The transcript uses ">>" markers but attribution is ambiguous in places. One speaker (apparently the guest, "Andrew") demos personal apps; the other demos company planning and recommends tools. **Uncertainty note**: without clearer speaker labels, per-person habits cannot be reliably separated. ## Uncertainty Statement - Some quoted fillers and phrasing repetition counts are approximate; the transcript's punctuation (repeated words like "have they have they," "we we we do") is reproduced as-is and may reflect transcription artifacts rather than actual stuttering. - Laughter and back-channel responses ("Yeah," "Totally," "Classic") suggest a rapport-building conversational style, but this is a single-session observation — no basis for claiming stable patterns across the speakers' broader output. # Position Record: "Agent-Native Apps: How to Build Apps After the End of Code" **Speakers** (identified from transcript evidence only): - **Andrew** — guest, builder of the demoed apps; states he oversees "30 CEOs" across businesses (addressed by name at the close) - **Host (Dan)** — Every.to co-founder; wrote his company's 2026 strategy, references COO "Brandon" and every.to --- ## 1. Andrew's "Deal Hunter" App Demo ### What it does - **Company analysis tool**: user names any company (public, private, or private credit deal), optionally uploads proprietary documents, and the app pulls company background, financials, leadership team info, and flags (e.g., lawsuits, financial crimes), then runs a valuation and generates a report and deal memo. - It did not retrieve all expected data in the live demo; Andrew noted: *"It's not doing evaluation because it doesn't know your numbers."* ### Technical stack (per Andrew) - **Claude API** as the core - **Tavily, Ampify, and Firecrawl** for financial data and checks ### Stated limitations — preserved verbatim > "It's again it's like not accurate... I'm still dealing with hallucinations and data being inaccurate and all that kind of stuff." - His workaround: export to PDF, show Claude, ask it to **fact-check everything**, and **reverse-engineer the right APIs** — but he acknowledges "it still... works" imperfectly. ## 2. The "Last 5%" Problem Andrew's central complaint about **vibe coding**: > "You can really get 95% of the way there in like 2 hours and then the final 5%. It just goes on forever." - He reports "going in circles over and over" on final polish/data accuracy after achieving the desired design. ## 3. Andrew's Personality App - ~30 minutes of questions → a roughly **100-page personality document** - Generates a prompt to build a **custom GPT/Claude/Grok project**; supports partner comparison, a **clinical PDF version for therapists**, and summary cards (communication style, strengths, best tasks), plus graphs of attachment style, interpersonal styles, motivations. - **Acknowledged bug**: every user eventually hits an issue that corrupts their results, wasting an hour of form-filling. - He emphasizes pride in design quality: it doesn't look like "some vibecoded... lovable" template. ## 4. Host's Diagnosis: Traditional Architecture The host argues the accuracy failures stem from architecture, not capability: - Recommends the **Compound Engineering plugin** ("keep going until this works" self-testing) and an **agent-native review skill**. - Claims: *"I would guess that this is actually a problem of this architecture being a little bit more traditional because that's just like how AIs think... You have to like prompt them to think this way."* - Andrew asks whether compound engineering requires narrow, feature-specific prompts; host responds it **can also be used** with broad multi-feature goals — Andrew's assumption isn't wrong, just not required. ### Andrew's counterpoint/qualification - **Scalability concern**: "This does need to scale... one analysis every 3 days." - His stated lesson: the importance of setting **guardrails, logging, and debugging/self-healing tools up front**. ## 5. Every's Agent-Native Company Planning (Host) - Host wrote the **2026 strategy** ("partially me and partially Claude"); COO Brandon built a **Notion agent** (an under-publicized feature) that **interviewed every department head** about last quarter, next quarter, and alignment with company strategy. - Output "more or less written by the Notion agent, more or less by Claude, but with a lot of back and forth from department heads"; the agent "asks tough questions" and helps check assumptions. - Triggering mechanism was mundane: *"a ping from Brandon being like, 'Please do this.'"* - Expanding use to **job applicants** and **editorial/copy editing**. ### Company positioning > "We're the only subscription you need to stay at the edge of AI" — the strategy theme is consolidating into "one company, one subscription" that is legible and serves specific people. ## 6. Discord Story-Pitching Agent (Host) - Claude Code set loose on the company's internal Discord (DMs included) to find interesting tools/workflows and **pitch newsletter stories**. - Dogfooding internally; host will only ship it if everyone says "this is the best." - **Accidental discovery**: the agent followed a link to a **Granola meeting recording** and surfaced an event the host would have had an opinion on but wouldn't otherwise have known about — he then intervened immediately. Host notes this happened by accident, and doing it deliberately would surface many more such events. ## 7. Meeting Recording & Change Management - Host's company uses **Granola** company-wide; experimenting with **Notion's equivalent** for centralization. No push-button whole-company solution exists to his knowledge; instead it's a **cultural expectation**: record meetings unless "super private." Team "bought into that." ### Andrew's position on change management > "It's impossible. It's like nailing jello to the wall... When you've got a cruise ship going a certain direction, it's very difficult to turn it." - Across his "30 CEOs," some love AI, some call it "overblown" — change management is the bottleneck. - He's excited about AI **custom-tailoring motivational messages per person** (e.g., revenue doubling for sales staff vs. design awards for designers), which he believes will make running companies "less and less painful." ## 8. "Read with Claude" / every.to Guide (Host) - A guide released that day includes a **"Read with Claude" button** — also works with ChatGPT. The prompt instructs the model to offer to go deeper on the reader's interests and, at a natural point, ask whether to remember to **search every.to** when AI topics arise. - Andrew corroborates the pattern: he did the same via **Atlas**, asking it to "summarize this for someone who's non-technical." ## 9. Tooling Discussion (both) - **Atlas**: both praise it. Andrew cites **native speed** as why he prefers ChatGPT over Claude desktop ("Spotify versus Apple Music problem"). Host uses Atlas agent mode for setup tasks like Streamyard/YouTube streaming configuration. - Host has an agent **building a Luma event** for a Vibe Code meetup in the background. - **Claude Code mobile**: Andrew criticizes it (must leave open, fiddly, no main-branch work). Host corrects one point: *you don't have to leave it open*, and it indeed can't work on main branch. - Recommendations from host: **Happy Claude** (happy.engineering, slash command syncs a session to phone), **Claude Island** (Claude sessions in the Mac dynamic island). - Host's personal workflow: fire off multiple Claude requests from bed/restaurant on his reader app, come home to **multiple PR branches**, merge into a testing branch, verify on phone. - Host does **not** sync his Claude directory via iCloud — he only has one computer. - **Opus Plan** (`/model` → "opus plan"): Sonnet for code execution, Opus for planning, dynamically switched — host finds it faster while retaining Opus planning intelligence. - **Sonnet 1M**: Andrew uses it for large codebase reviews. ## 10. Closing Sentiment - Andrew expresses envy of entrepreneurs starting today and of the host's AI-native leadership team; frames AI as removing his bottleneck of "having to convince other people": > "It's like I'm riding a bike and I was walking before." > "I have a helicopter, like a flying drone that flies me wherever I want in 10 seconds." - Host plans to try the **agent-native architecture** on the deal app and report back. --- ## Unresolved Tensions - Andrew's demo apps remain **unreliable in practice** (hallucinations, data gaps, corruption bug) despite rapid development — the host's "agent-native architecture" claim is offered as a hypothesis, not yet tested on Andrew's apps. - Andrew's **scalability requirement** vs. agent-native approaches is raised but not resolved. - Andrew's organizational reality (skeptical CEOs, change management as "nailing jello to the wall") contrasts with the host's small, AI-native team — acknowledged by both, not reconciled. # Speaker Attribution Record — "Agent-Native Apps: How to Build Apps After the End of Code" ## Speakers - **Speaker A (Guest, "Andrew")** — Identified by name in the closing ("Andrew, uh, thank you so much for doing this with me"). Demonstrates self-built apps: a company-analysis "deal hunter" app and a personality assessment app. Describes his organization as having "30 CEOs" across multiple businesses, implying a portfolio-company structure. - **Speaker B (Host/Creator)** — Runs the session, thanks the guest, presents his company's 2026 planning in Notion, references his COO **Brandon** by name, and mentions a guide released "today" with a "read with Claude" button. Company uses Discord internally and is "20-ish people." Uncertainty note: first names "Andrew" and "Brandon" are explicit; no full identities or channel affiliations are stated in the transcript. No curated metadata was supplied. ## Referenced Tools, Products, and Services Cited by **Speaker A (guest)**: - **Anthropic Claude API / Claude Code** (mobile app), **ChatGPT**, **Grok** - **Tavily** and **Ampify** ("Tavali"/"Ampify" in transcript) — financial data pulls - **Firecrawl** — data crawling - **Lovable** — cited as a template-y vibe-coding tool - **Fireflies** / **Otter** — meeting recording agents - **Granola** — meeting notes - **Playwright** ("playright") — testing bugs - **Happy Claude** (happy.engineering) — mobile Claude Code sync - **Claude Island** — Claude sessions in Mac dynamic island - **Opus Plan** model mode; **Sonnet 1 million** context - **Atlas** — praised for native speed/agent mode Cited by **Speaker B (host)**: - **Claude**, **ChatGPT**, **Notion** (incl. an unpublished Notion agent feature), **Granola**, **Claude Code**, **Loom**, **StreamYard**, **YouTube**, **iCloud** - **Compound Engineering plugin** and **"agent native review" skill** — recommended to the guest ## Third-Party Claims & Organizational References - Host's company: 2026 strategy doc "partially me and partially Claude"; Notion agent interviewed departments; company is "the only subscription you need to stay at the edge of AI"; ~20 people. - Guest's companies: ~30 CEOs; change management friction described as "nailing jello to the wall." - Guest's apps include reported inaccuracies/hallucinations, per his own statement. ## Promotional Segments - **Host promotes his own guide** ("we released this guide today") featuring a **"read with Claude" button** and a prompt instructing the AI to offer to search "every.TO" — the domain suggests promotion of every.to; treated as promotional. - Host promotes **Compound Engineering plugin** and related skill ("we'll put it in the chat"). - Guest expresses enthusiasm for **Atlas** ("Atlas is awesome") — unprompted praise, not flagged as paid. ## Unavailable Evidence Channels - No curated metadata, timestamps, channel identity, subscriber data, or comment availability were provided. - Speaker identities beyond first names "Andrew" and "Brandon" are unverified. Let me show my screen. Okay. Can you see this? Uh, yep, we can. \>\> Okay. So, it's pretty cool. So, basically, I built this as like an initial screen for businesses we're looking at. So, \>\> can click analyze a company. I can say, you know, private, public, or private credit deal. Um, now you can name any company. So, let's name um is it INC? Every Inc. \>\> Yeah, Every Media Inc. \>\> Every Media Inc. Well, let's see how it does. Every media New York and let's see. And then basically, you can upload any like proprietary documents you have or whatever. \>\> And then you \[clears throat\] just say analyze. Now, it's going to go out and get um company background, any financials it can find. It's going to look at the leadership team and look at like have any of you \>\> Yeah, I'm I'm waiting for the flags to come up here, but it's like, \[laughter\] you know, have they have they like been sued by a million people? Have they committed financial crimes? Blah blah blah. And then um it runs evaluation, generates a report, and then we can develop a deal memo based on it. Um let's let this run. It might be broken, but um but anyway, it's been not helpful yet, but you can see where it's going um rapidly. I'm still dealing with hallucinations and data being inaccurate and all that kind of stuff. Oh, yeah. Here we go. \>\> Very cool. What What is What is running this? Is this opus in the background or what is the \>\> this is um yeah it's all it's all using the cloud API and then I use um Tavali and Ampify to pull financial data um and like do the actual checks and fire crawl a little bit. Um it's again it's like not accurate but what's shocking is just how far I was able to get in a very brief period of time. I think the most infuriating thing about vibe coding is that you can really get 95% of the way there in like 2 hours and then the final 5%. It just it goes on forever. \>\> Classic. \>\> It's insane. Like on my um let me open my personality app while we wait for this. \>\> On my personality app, I've had this like insane um bug where everyone that tries to use it at some point has an issue and then all their results are corrupted. And so it's like they just spend an hour fill out forms and boom, you know, here they are, right? It's like, \>\> so this is this is the other app while we wait for the other one. This is the personality app. Um, and so basically \>\> I only see um the deal hunter screen. \>\> Oh, here. Let me fix that. So, so basically the once you fill out you do about 30 minutes of questions, it writes you this massive document like literally like I think this is a 100 pages of like all every aspect of your personality in life and then it also builds this whole thing. So basically once you have all your results it has this thing where that lets you create a custom AI. So you can create a chat GBT GPT or a project or a cloud project or Grock and it gives you a prompt that basically breaks down all of your different um all of your different scoring and then you can invite your partner and compare with them. You can send it to your therapist. So it'll actually create like a clinical version of the PDF and all your issues and then it also creates these cards. So like one is like what's the best way to work with me? Um what's my communication style? What are my strengths? What are the best tasks for me etc. And then we've got all these cool graphs that show you know your attachment style um you know different interpersonal styles, motivations. Anyway, it's just like it's so cool. It's so useful. Um I found it so useful for me and I'm like so excited to show everyone. But I'm just mostly I'm just astounded by the level of quality of the design and how quickly I was able to build it. \>\> It's just so comprehensive. Like there's so much to it and it looks it looks real. Like it doesn't look like some \>\> No, it's not like some vibecoded like you can tell when someone's used like a lovable \>\> or used lovable and there's just kind of like a a template or whatever. So So here's the deal hunter. It's come back on every \>\> I want to know what my m what my multiple is. \>\> Yeah. So, let's see. It's saying highly specul speculative acquisition opportunity. \[laughter\] \>\> It's saying that it's saying that due to lack of publicly available information. \>\> Um, so here's like a breakdown of the company, how many employees. \>\> Um, it should get your name. I don't know why \>\> you got to change your uh your algorithm a little bit. \[laughter\] \>\> I don't know why it didn't get that. Um, and then it's not doing evaluation because it doesn't know your numbers or whatever. But anyway, all this would be filled out if I uploaded financials and stuff. \>\> Does it have a regular web search tool? \>\> Um, it does. Um, I don't know why. Yeah, this is this is the problem. So, I got to this point of like the design looking the way I want and everything. And then the last 5% has me been it's been me going in circles um over and over and over again. um trying to get this stuff to to actually get the results right. And so what I'll do is I'll like export it as a PDF, show Claude, tell it to fact check everything, and then reverse engineer to use the right APIs to get the right information, but it still, you know, works. \>\> I have a couple of suggestions for you. Um, if you install the Compound Engineering plugin, um, one thing that you can do, and we'll I'll I think you have it, but we'll we'll put it in the chat, but one thing you can do is you can have it test itself \>\> and just say like just keep going until this works, which I love because then you just walk away and you're just like you just watch it working and it's just like it's just going until it works. But I think we we also have this skill that's um agent native review which will review your app for like how well it conforms to agent native principles. And I would guess if you made it like slightly more agent native maybe not fully but slightly it would fix a lot of these issues. It would like be smart enough to be like yeah I know I have to fill out this section. I'm going to do a bunch of web searching until I find out all the stuff because there's a lot of public information public information about every on the internet. Um, so I would guess that this is actually a problem of um this architecture being a little bit more traditional because that's just like how AIS think. They're not they're not trained to think this way. You have to like prompt them to think this way. \>\> Well, this does need to scale, right? Like we're going to do one analysis every like 3 days or something. Um, I totally I think like where I'm at with vibe coding now is like I am starting to realize how important it is to set up the guard rails in the beginning to have proper logging and give the AI all the tools it needs to actually debug and selfheal and all that. And I'm really I'm curious like compound engineering the only thing about it that I was slightly confused by is it seems like you need to be very specific about what you want to build. like go build this one feature, not like what I've been doing has been like, you know, hey, here's like the end goal. Go build like five features at once. Am I wrong? \>\> Yeah, you don't need to do it. You can you can use it in the way that you're you're you're envisioning for sure. \>\> I've used I've used it for a few um code reviews and it's been amazing. Although I had a few bugs with the playright, but I think that's been resolved. Now, \>\> one thing I want to show you, and I don't know how much time you have left, but one thing I want to show you that I think you're going to find fascinating, and this is on agent native architectures, but it's actually agent native ways of building a company. Um, so we we used Notion to do our all of our 2026 planning and our COO, Brandon, just like went super hard on using agents to help all of us um plan for each department in the company, like make our 2026 plan and then make sure it conformed to the overall goals for the company. So, see if I can share my screen again. Dude, I just got to say you're so lucky that you have so many people at the top across all your your whole business that are all like AI native. I really like our pro our problem is like you know we've got you know 30 CEOs um you know some of them love AI some of them they're like oh it's all overblown whatever it's very difficult to do the change management across so many businesses. \>\> It really is. I'm super lucky. Um, and I that's why I love like our whole business is to is to do this because it filters for people who are like just all in on trying to do it well. So this is our 2026 strategy doc. And basically Brandon was like, "Okay, Dan, I just want you to write our 2026 strategy like the top level strategy." So I did that and I did this like partially me and partially Claude. And it's like, "Okay, just super clear. You know, we're the only subscription you need to say at the edge of AI. How do we do it? who do we do it for? All that kind of stuff. What's our theme? You know, I think we've been super we've we do a lot of different things and we just need to like pull everything down into like one company, one subscription that's like really legible and that everyone can understand and and provides a lot of value for the specific people that we're serving. So, any I wrote that 2026 strategy and then what he did is he created a notion agent, which is like a new feature that they haven't really publicized that much yet. um and it interviewed everyone in every department about um how they did last quarter and then what they're going to do this quarter and then how it um how it uh levers up into the overall company strategy. And so all of this is basically written more or less by the notion agent, more or less by Claude, but like with a lot of back and forth from all the department heads and it'll like it asks tough questions. It helps you check assumptions. it. Um, \>\> is that is that an email or what are they getting? Like a a message in notion \>\> uh to to do this. \>\> Uh, it's just like a ping from Brandon being like, "Please do this." \[laughter\] Um, and so it just made everything about how we, you know, we're now we're 20-ish people or whatever. It just made everything about how we run the company so much easier and so much better. And um, we're we're like just scratching the surface. Now we're doing this for people who apply for jobs and um we're starting to do this for uh we have a bunch of stuff in our in our editorial like how we do copy editing and how we do other kinds of edits to make sure that pieces we're writing are like you know top grade. It's it's really cool. It's really interesting. \>\> It's crazy. I'm I'm so like I'm the kind of person like I I love systems like this, but what I don't like is change management and I don't like trying to get all the people to use the system, but I feel like this just kind of makes it so incredibly easy that it feeds through the entire business. Um I'm so excited for I feel so jealous of like entrepreneurs starting today. Like this stuff like this just it makes the idea that you know you could build an agent that basically goes through every single imagine you have a thousand person organization. It goes through every single person and gets their feedback on the strategy boils up all the information to the CEO you know just allows the CEO to basically talk to everyone in their own customized way and get feedback from everyone. Like I just think it's totally unheard of and insane and previously would have required like this huge HR team. Totally. And and like I had this experience the other day where I I've been experimenting with okay if we started every today and we were doing a newsletter today. What would that look like? Because I don't I don't think it would be the thing that we have. And so one of my ideas was okay I'm going to make cloud code. Eventually I'll make this a whole agent, but right now it's just cloud code and it's going to go through all of our uh like our whole Discord. We use Discord for our internal company operations. It's going to go through our Discord and just find all the interesting \[ \_\_ \] that we're talking about and then it's going to pitch me stories based on whatever we're talking about because there's always like a new tool or a new workflow or whatever. And so I just set it off. I was like, "Go log in my Discord, go through my DMs, go through everything and pitch me some ideas." And a it came up with some really good stuff. And like I then sent out um we're just we're dog fooding it internally because I just I only want to do it if everyone ever is like this is the \[ \_\_ \] best. Like I read this every day, so we're starting to dog food internally. But one of the things that happened is it was going through our Discord and it found a link to a granola meeting that had happened that day. And then it went and read the granola meeting and then it told me about something that had happened in that meeting that I would have had an opinion on that I would have had no idea that it had happened. And I was immediately able to just be like, "Hey, like I just saw this happen in this meeting. Like do this." Um, and that's crazy. and and that's that that happened by accident, but if you were doing it on purpose, I'm sure there's like plenty of stuff that's happening in plenty of meetings that I want to know about that I I now can know about. \>\> Do you record every single company meeting? \>\> Not every single, but we're moving in in that direction for sure. Um, for a long time \>\> like one problem I have is like I have a um AI agent. I use Fireflies or Otter depending on the day and it auto logs into all the zooms but sometimes it doesn't get in or you know whatever. have you figured out an architecture whereby you can basically plug into every single meeting because what I would love is to be like what's happening at tiny like what's at the latest with that deal and not have to bother someone and just know \>\> yeah um I just I just use granola and like the whole team basically uses granola we're starting to switch experiment with switching to notion and because they have a very similar thing but the difference is it's in notion which is you can connect to everything so it's like it's it's nicer to have it I think centralized um but yeah it's not Uh I I don't know of a push button thing where it would automatically add it to a com like entire company's meetings, but it's just culturally a thing that we're setting as an expectation that if you're in a meeting, if it's not like a super private one, you should record it so that we can figure out we can use it later and I think everyone's bought into that. But I imagine I think you're running into which I'm really curious about is like change management at more mature organizations. Like how are you guys even doing that? \>\> Well, it's it's it's impossible. It's like nailing jello- to the wall, right? When you've got a cruise ship going a certain direction, it's very difficult to turn it. And and I really believe that what it requires is one-on ones with every single person, at least the key management. And that's just not tenable if you're super busy or you have so many businesses or whatever. Like I am fascinated by the idea of being able to actually custom tailor messages based on the person, right? There's certain people that are motivated by different things. And so if you go to one person, you say, you know, every in 2016, we're going to double our revenue. That might be really exciting to someone in sales, but one of your designers needs to hear we're going to release the greatest projects, you know, or the greatest products ever, and we're going to win an Apple design award or whatever. And I think um that is what I'm super excited about. Like running a company is going to become less and less painful over time with this stuff. \>\> Totally. This is actually one of the things I did um when so we released this guide today and one of the things I like about this guide is um it has a button that says read with Claude, right? And if you click it, it'll go into your Claude. And if my internet is not being slow, can you still hear me? Yep. Okay. Um let's see if it'll let's see if it'll work with um GBT. So for chat GBT um and it's the same thing with cloud. Like it it does the same thing. It'll say, "Hey, I got something cool for you. Curious what you make of this." Puts in the link. It's a framework for building agent native apps. I just press send. And what it does is it says then offer to go deeper on what's most interesting or relevant to me and at some natural point in the conversation ask if you like me to remember to search every.TO when AI related topics come up. And what I love about this as like a test is it it automatically takes the article and puts it into a form that is um much more default useful for uh whoever's reading it. Like this is a long big thing. Well, that's actually exactly what I did. I went in Atlas and I said, "Summarize this for someone who's non-technical." And then it broke it all down. \>\> Exactly. \>\> By the way, I love Atlas is awesome. It's It's a great testament to um just native speed as well. \>\> Like how fast it is and I think like so much of why I use Claude versus or so so much of why I use Chat GT over Claude desktop is just speed. It's just because it's native. They really It's like the Spotify versus Apple Music problem. They really need to think about that. \>\> Yeah, I use um I use Atlas all the time because I love the agent mode. Like yes, setting up the stream for example. It's like okay, I got to like figure out I got to sign up for Streamyard and then I got to figure out how to connect to YouTube and there's like 50 different keys and settings in all these different places and I just said Atlas figure it out and it more or less did that. Um it's so much better. \>\> Totally. I'm actually um I have an agent going right now because I want to do a Vibe Code meetup. So, I've got it building a Loom event in the background here right now. I was actually just prompting it to keep going, \[laughter\] but it's it's like the I just feel like pinch me like this is all just so insane. And it's for anyone that's high agency. Like I've always thought like I'm very high agency, but I what I don't have time for is people, right? Not in a negative way. love people, you know, for friends or whatever, but when it comes to actually wanting to do things, having to convince other people to do the thing and having to have this chain of people to accomplish anything was just so infuriating. And I'm finding like it's like I've I'm riding a bike and I was walking before. It's just like insane. \>\> I'm I'm with you. \>\> I have a I have a helicopter, like a a flying drone that flies me wherever I want in 10 seconds. Honestly, it's insane. \[laughter\] \>\> I totally agree. Um, I think we're we're probably out about time. Andrew, uh, thank you so much for doing this with me. It's so fun to to get to hang out with you. \>\> Yeah, that was sweet, dude. \>\> Oh, I have one final question. How are you doing, um, Claude? So, I tried using the Cloud Code mobile app. It kind of sucks because it like just you kind of have to leave it open and a little fiddly and then you have to do a merge for you can't work on your main branch, right? \>\> You you it does not work on the main branch, but you don't have to leave it open. Which are you using just a regular cloud app? \>\> Yeah, I guess I guess I thought I had to leave it open, but are you doing SSH or using any tools just to like No, there's one I just found called um Happy Claude, which is quite good. So, what you do is you when you're on a thread and you're going to leave to go to your computer here, I'll I'll put this in the chat. Um \>\> it's happy.engineering. And what it what you do is you just write like a slash command like slashmobile or something and then it syncs with your phone and you can just leave and then you have a fully direct line into that terminal session. Yes, \>\> that's really cool. \>\> The other one \>\> Yeah, \>\> the other one I was going to recommend is uh Claude Island. So what it does is it puts all your um Claude sessions into your uh dynamic island on your Mac and it's just an easy way to like get them going. Yeah, \>\> that's really nice. Yeah, I I've been using the the cloud code, like the mobile app inside of cloud. I think it I didn't like it when it first came out, but I think it's actually much it's much better now. It's not the same thing at all as like doing it locally, but um I think the experience is actually pretty nice where basically what I do is I just, you know, I'm using these apps all the time. I have a reader app that I made for for myself, for example, which is the same underlying agent native architecture. and I'll just be reading in bed or like at a restaurant or whatever and I'll be using the app and then I'll find all these like little things I want to be different and I just like fire off a fire off a a cloud for this and for that and for that and then I get home and I have like four different PR branches waiting for me and I just merge them all into one testing branch, put it on my phone and then keep going and then merge it later if it if everything works. It's like a really good workflow. \>\> So you're and you keep your cloud directory in iCloud is that correct? So it's synced across all your computers. No repos. \>\> I don't do that. You don't do that. I mean, I probably should if I was like being really I don't I actually don't even have another computer. I just have this one. So, right. \>\> Um I don't have a I don't have a reason to do that. But uh just Claude itself has it. It just the Claude app syncs all the conversations that you're having, not the ones on your computer. \>\> Do you know about Opus plan? \>\> No. So there's a there's a model that um if you go slashmod and cloud code and then you write opus plan it's sonnet for the code review \>\> then opus for the planning and so it'll dynamically switch between them for their highest and best use I find it's much faster and you're still getting the intelligence of opus for your planning. \>\> Interesting. \>\> And then also do you know about sonnet 1 million? \>\> Yes. \>\> Yes. What do you what do you use that spec in a in a specific \>\> I use it if I've got like a really big codebase review or something like that? \>\> Yeah. Sweet, \>\> dude. Amazing. Okay, I'm really excited about this agent native architecture. I'm going to try it out and then get back to you. \>\> Check it out. Yeah, if you like I know you're you're doing the journal app or um the deal app. If if you if you are able or want to like reorganize it with that, I'm really curious how it goes. \>\> Sweet. Okay, dude. That was fun. Have a good day. Bye.