PodcastsHear the voice. See the shape of the thought.

Figma developer PM Matt Colyer has been building his own AI agents for two years and is buying more software subscriptions than ever — not fewer. He and Every CEO Dan Shipper work through why the "SaaS apocalypse" narrative gets the economics backward, how AI needs to escape the tyranny of the text box to unlock genuinely creative design work, and why the coming year's challenge isn't generation but review: humans are now the bottleneck in a world where agents can ship faster than anyone can evaluate what they made. ## [00:00] AI will create a billion developers This exchange, taken from later in the interview, opens the episode: Matt argues that the number of developers worldwide — roughly 25–40 million a decade ago — is heading toward a billion. That demographic explosion, not AI replacing software, is what makes the SaaS market a "gold mine." Figma and most established SaaS businesses are, in his view, excited rather than threatened. > *"If you're in that space, like, it means it's a gold mine, right?"* ## [01:03] Introduction Dan Shipper frames the conversation: he recently bought Figma stock after noticing the "SaaS apocalypse" discourse, and he wants to know how a company that pre-dates AI is navigating a world where agents can now operate inside your product. Matt, as the director managing Figma's developer products, is the right person to ask. > *"There are all these people who are like, 'Oh, I don't have to use Figma anymore.' You guys just launched an agent in your product. You also have Figma MCP."* ## [02:15] Why the SaaSpocalypse narrative has it backwards Matt's counter-argument runs on two tracks. First, the democratization of software creation massively expands the addressable market — more software being built means more demand for the tools, infrastructure, and services that support it. Second, vibe-coding your own app sounds liberating until you're dealing with SMTP upgrades at midnight. He built his own email agent two years ago and watched it get rickety; these days he pays someone else to run agents for him rather than maintain the plumbing himself. > *"I'm buying more software these days than I ever did before, because I'm like, 'You know what? That tool seems cool. I'm just going to pay somebody else to run my agent for me.'"* ## [05:27] Matt's email agent origin story The origin was unglamorous: three kids in three schools, relentless PTO emails, and the humiliation of missing spirit day. Matt wired up a Python script to grab his inbox and paste it to an LLM — the whole thing was rickety and sometimes the replies didn't work, but the core loop worked. He then added a memory system and a daily summary pushed to him proactively, which he flags as the real unlock: instead of having to open a tool and ask, it just showed up. Dan mirrors this with his own Codex-based inbox workflow, now four weeks into inbox zero. The two also land on voice as an underrated interface — Matt uses Loom recordings because it feels less weird than talking to a blank screen. > *"The unlock for me was like instead of having to go to a tool and ask for the thing, it was just like it would show up."* ## [13:21] Divergent vs. convergent design thinking Chat-based AI is inherently linear — you iterate on one design thread. Matt's argument is that great design has a diamond shape: first you diverge (generate many directions), then you converge (pick the best). Figma's on-canvas agent is a first attempt to break out of the text-box constraint. On the canvas, an agent can spawn a grid of frames — grayscale, sepia, with different type — and then a separate convergent agent can cluster them and recommend which direction to pursue. Command-line agents can't do this kind of spatial, parallel exploration; that's what the canvas unlocks. > *"Text boxes are super limiting — it's very much like a linear 'well this and then that.' If we get to the canvas, the agents allow you to do divergent thinking."* ## [17:39] Figma's MCP server MCP gives third-party agents (Cursor, Windsurf, Claude Code) a standard interface into Figma. Two flows: code-to-design — fire up a dev server, ask the agent to screenshot a live page and pull it into a Figma canvas — and design-to-code via "Get Design Context," which wraps component properties and design library guidelines into an agent prompt that then creates a branch, writes the code, and posts a screenshot to the PR. Both flows remove the manual copy-paste drudgery that used to live between the design file and the codebase. > *"You pull up your codebase, fire up the MCP server, and ask it, 'Hey, can you go to this page and copy it into Figma canvas?' And it will actually do it. That's a little bit mind-blowing."* ## [19:45] Why design agents need personalization Generic agents produce generic output. For Figma, the difference between an okay agent and one people actually love is whether it understands the design system — the components, the spacing rules, the naming conventions. Without that personalization layer, generated designs aren't usable. Matt draws a parallel to the memory systems in chat agents: in Figma's case, the design library is the memory. He also hints at proactive agent work Figma is cooking internally, framing the core problem as maintaining design values at a pace agents can generate. > *"The thing that really differentiates an okay agent from one that people really love is the personalization aspect. For Figma's version of that, it's the design system."* ## [22:09] Every problem is a context problem Matt describes a Figma product operations team that realized every recurring PM task — onboarding docs, project tracking, team introductions — was a context problem in disguise. They built "PMOS": a local SQLite org chart wired to Asana, Slack, and GitHub, then layered Claude Code skills on top. When a new team member joins, the system walks the org chart, reads the last 30 days of Slack channels, checks the Asana board, and produces an uncannily good onboarding file. Dan points out that Claude Code's power comes from the same insight: instead of an always-on cloud agent you have to manually wire to everything, it's an agent that already has access to everything on the user's machine. > *"One of the unlocks to me about AI is like you kind of realize every problem becomes a context problem. The work becomes about framing the problem with the right set of information."* ## [25:12] Apple and Google as the reigning kings of context Matt has been waiting for Apple Intelligence to deliver on its WWDC promise — phones hold all the personal data; an always-on, actually-smart Siri should be the obvious product. It hasn't arrived. He's watching Google's rumored "Spark" agent (always-on, connected to all Google content) with similar anticipation. Dan's take: Apple wins regardless because everyone runs AI on Mac hardware, giving them time to catch up. Matt adds that Apple's privacy-first positioning is a genuine strategic asset, not just PR. > *"Even being late to the game, they are still the king of context. And I think that's what's been interesting to watch about Google I/O this year — seemingly Google has also kind of woken up to that."* ## [28:18] Why review is the new bottleneck Generation is no longer the hard part. Agents are cheap, capable, and available; the problem is that humans are now inundated with net-new content they need to evaluate and approve. Matt frames "review" as the coming year's core design challenge: how do you scale a human value system — what good looks like, what fits your brand — at the pace agents can ship? The format is still unsettled: video walkthroughs, screenshots, a trusted review agent. He closes with a thought on careers: fundamentals still matter (you need to know what long division is even if you use a calculator), and the people who will thrive are the curious ones who ask how something is put together rather than just accepting the output. > *"We have agents that are capable of producing all this stuff, they're available enough, they're cheap enough. We're just being inundated with new content. The bottleneck is now: how do we scale our value system to evaluate it?"* ## Entities - **Matt Colyer** (Person): Director of Product Management for Developers at Figma; has been building personal AI agents for two years; longtime developer tools practitioner. - **Dan Shipper** (Person): Co-founder and CEO of Every; host of the "AI & I" podcast; active AI agent practitioner (inbox zero via Codex). - **Figma** (Organization): Design and prototyping platform; launched an on-canvas agent and an MCP server; central example in the SaaS-in-the-AI-era discussion. - **SaaSpocalypse / SaaS Apocalypse** (Concept): The narrative that AI will make SaaS software obsolete; both guests argue the opposite — AI expands the developer population and demand for SaaS. - **Diamond-shaped design thinking** (Concept): Divergent phase (generate many options) followed by convergent phase (select the best); Colyer argues current chat-based AI only supports linear/convergent work. - **MCP (Model Context Protocol)** (Concept): Standard interface for third-party agents to connect to tools like Figma; enables code-to-design and design-to-code workflows. - **Figma MCP Server** (Software): Figma's implementation of MCP; supports live page screenshot-to-canvas import and "Get Design Context" design-to-code export. - **Claude Code** (Software): Anthropic's coding agent; referenced as an example of an agent with full local file system context; used by Dan Shipper for inbox management. - **Every** (Organization): AI-focused media and software company; Dan Shipper is co-founder/CEO; runs the "AI & I" podcast series. - **Proactive agents** (Concept): Agents that push summaries or actions to users without being asked; Matt identifies the proactive daily email summary as the unlock that made his agent genuinely useful. - **Review bottleneck** (Concept): The emerging constraint in AI-assisted work where generation is fast but human evaluation/approval capacity is the limiting factor.

Dan Shipper, CEO of Every, delivers a day-zero vibe check on Opus 4.8, arguing Anthropic could have called it Opus 5. The model jumps 30 points past Opus 4.7 on Every's Senior Engineer benchmark, edges out GPT-5.5, tops their internal writing tests at 79.6 vs. 73, and is the first model to produce a genuinely good one-shot slide deck. Two catches temper the enthusiasm: performance degrades sharply below "extra high" reasoning, and the Claude desktop app remains cluttered compared to Codex. ## [00:00] What is Every Every is a 30-person applied AI lab for the future of work—part media outlet, part product studio. Dan opens by explaining the subscription (writing, courses, AI-built tools all in one place at every.to) before rolling into the Opus 4.8 assessment. The plug is brief and context-setting: the team has had beta access for a week, and the rest of the video is what they found. > *"Every is the only subscription you need to stay at the edge of AI."* ## [01:07] Anthropic Is Back: The Headline Case for Opus 4.8 Dan had largely abandoned Claude after Opus 4.7—slow, hard to love, and outpaced by Codex and GPT-5.5 in day-to-day use. Even the most loyal Claude users at Every had started routing work elsewhere. Opus 4.8 breaks that pattern: it scores 63 on Every's Senior Engineer benchmark (30 points above Opus 4.7, one point above GPT-5.5), tops their writing tests, and produced the first one-shot slide deck Dan has called genuinely good. Kieran Klaassen, Every's GM, called it "the most human model he's worked with." The one persistent friction is the Claude desktop app itself. Codex is fast, focused, and ships a clean harness; the Claude app still feels like a product built by three separate teams—chat tab, code tab, co-work tab, each with its own feel. Dan is now splitting time between both apps, which he was not doing before. > *"But honestly, they could have called it Opus 5 cuz this is a really great model."* ## [05:02] Reach Test: Paradigm Shift Ratings from the Every Team Every's reach test asks one question: do you actually open this model when work gets hard? Dan rates Opus 4.8 gold/green—paradigm-shift quality, docked one notch because the Claude app harness is only "okayish to pretty good." Kieran, who runs 50 agents a day, gives a straight gold paradigm-shift, one of the rarest grades the team has assigned. Katie Parrot, a senior staff writer and historical Claude fan, lands at green, splitting her work between Opus 4.8 and Codex. > *"It's very rare to give a paradigm shift grade to a model. So I would pay attention to this."* ## [06:32] Benchmarks: Coding and Writing Numbers On coding, Opus 4.8 hits 63 on the Senior Engineer benchmark—the test feeds the model a vibe-coded codebase and asks it to rewrite from first principles, then scores against two human senior engineers who completed the same rewrite (typically scoring in the 80s–90s). GPT-5.5 sits at 62. On Kieran's LFGbench (real-world tasks: SaaS build, e-commerce site, 3D game landscape), the model writes readable code that bridges technical competence and creativity—the "cozy island" 3D scene is notably richer and more vibrant than GPT-5.5's output. On writing, Opus 4.8 scores 79.6 out of 100 on Every's internal benchmark (intro writing, promo emails, mid-piece paragraphs); GPT-5.5 scores 73. The gap is mainly in AI tells: at high and extra-high reasoning settings, Opus 4.8 produces prose that sounds less like a model. It matches a writer's voice from a single paragraph of context better than any other model Dan has tested. > *"Opus 4.8 scores a 79.6 out of 100 on the writing benchmark. GPT 5.5 is 73."* ## [08:57] Emotional Intelligence, Knowledge Work, and the Verdict Dan uses the model for interpersonal and management work—talking through decisions, pressure-testing his own framing. Opus 4.8's thinking traces show it genuinely cycling through permutations before responding, which makes it feel less like a sycophant and more like a useful counterpart. On knowledge work, it's versatile: code and writing coexist cleanly in a single thread, and the slide deck result is the first one-shot deck Dan would actually send to someone. The verdict: if you're a Claude fan, this model delivers. If Codex converted you, add Opus 4.8 as a parallel tool for writing and knowledge work—it's worth the context switch. The harness gap is real, but the model itself is a banger. > *"If you've been converted to Codex, I highly recommend you at least add it as part of your arsenal."* ## Entities - **Dan Shipper** (Person): Co-founder and CEO of Every; presenter and primary evaluator of Opus 4.8. - **Kieran Klaassen** (Person): GM of Kora at Every; gave Opus 4.8 a straight gold paradigm-shift rating on the reach test. - **Katie Parrot** (Person): Senior staff writer at Every; rated Opus 4.8 green, split between it and Codex. - **Every** (Organization): Applied AI lab and media subscription company focused on AI for the future of work. - **Anthropic** (Organization): Developer of Claude and Opus 4.8. - **Opus 4.8** (Software): Anthropic's latest Claude model; subject of the vibe check. - **GPT-5.5** (Software): OpenAI model used as the primary performance comparison across all benchmarks. - **Codex** (Software): OpenAI coding agent; praised for its clean desktop harness and used as the daily-driver counterpoint to Claude. - **Senior Engineer Benchmark** (Concept): Every's proprietary coding benchmark—rewrites a vibe-coded codebase from first principles and scores against human engineers. - **LFGbench** (Concept): Kieran Klaassen's real-world coding benchmark covering SaaS, e-commerce, and 3D scene generation tasks.

Dan Shipper's Every has grown from four people to thirty since GPT-3, runs agents in nearly every workflow, and is still hiring. In a format flip for the *AI & I* show, COO Brandon Gell interviews Dan about his 8,000-word essay "After Automation," which argues that rising AI capability creates more demand for human judgment, not less. The core mechanism: AI makes yesterday's expert competence cheap and ubiquitous, which floods every domain with output that's close but not quite right — and that gap drives more work for the humans who can close it. ## [00:00] AI does it, then asks what's next This exchange from later in the interview captures the central tension of the episode. Brandon describes the archetypal AI moment — you prompt it, it blows your mind, you feel obsolete — and then it stalls and asks, "What should I do next?" Dan counters with the line that anchors the whole argument: "The further away an agent gets from a human, the less valuable it is." Both clips come from the main conversation (around 00:11 and 00:35 respectively), surfaced here to frame what follows. > *"The further away an agent gets from a human, the less valuable it is."* ## [00:51] Introduction Brandon sets up the format flip: he's interviewing Dan, not the other way around, and will push back on Dan's thesis. Dan explains the piece's origin — sitting inside one of the most agent-native companies in existence, watching headcount grow alongside automation, and feeling a disconnect from the mainstream narrative that AI is eliminating jobs. The ClickUp CEO's recent tweet (firing a large portion of staff and citing AI) drops into the conversation as the first stress test for Dan's argument: does "After Automation" hold for a 10,000-person mature company, not just an early-adopter shop like Every? > *"If you swing a stick around in our Slack, you're as likely to hit a human as you are an agent."* ## [05:51] The AI paradox: more automation, more human work Dan walks through the core argument. AI is trained on all prior outputs, so it can deliver "yesterday's expert competence" cheaply and to anyone. That democratizes output — ops people merge pull requests, non-engineers ship features — but the output is uniformly *close, not right*. It's not calibrated to the live situation. So you get a glut of near-correct work that devalues on its own, while simultaneously creating more demand for experts who can take that near-correct work across the finish line. Brandon adds the inside-Every version: PRs that look plausible until a senior engineer looks under the hood. > *"You sort of flood the zone with tons of stuff that's like close, but not quite right."* ## [10:00] How AI makes yesterday's expert competence cheap Dan extends the argument to the benchmark objection: yes, models improve exponentially, but once a benchmark saturates you can always unsaturate it by reframing the problem slightly. The deeper issue is that humans carry a layer of tacit, unarticulated competence that evades clean specification — and anything you *can* articulate, a model can hill-climb on. Every's experience bears this out: Kieran built a complete inbox feature end-to-end in a month or two, which was "completely impossible" before. But the value came from an expert knowing *what* to build and steering every step. > *"There's actually a lot of stuff that you do that can't be articulated in a clean frame."* ## [18:00] AI can act autonomously but it does not have agency Brandon draws the autonomy/agency line: AI agents are getting very good at executing open-ended tasks without hand-holding, but that is categorically different from *agency* — the self-motivated, playful, "I just want to do this because I'm into it" drive that even a toddler has. Dan agrees there's no economic incentive to build that: if you're at your desk and the agent says "nah, I'm playing," that's a product failure. The entire industry's incentive structure pushes toward compliance and corrigibility, which is exactly what keeps humans in the loop. > *"Agent means something that acts on behalf of someone else. That is very different from having agency, which is what even the smallest child has."* ## [20:39] Why Dan is all in on AGI Brandon proposes a one-word-answer test: do you think AGI will happen? Dan: yes. Is that a good thing? Dan: yes. His AGI definition — any agent that makes economic sense to run continuously, actively generating tokens and completing tasks without re-prompting — is precise enough to be testable. His reasoning: even a truly autonomous system will have been built to serve human goals; if it weren't, we wouldn't build it. Brandon's worry is that once continuous agents are economically rational, the mass-layoff argument becomes coherent. > *"Any agent that you never turn off — that it makes economic sense to keep running all the time, actively doing tasks without you ever having to re-prompt it."* ## [21:57] AI layoffs are a lie Dan and Brandon dissect the ClickUp case — a CEO who publicly fired a large portion of his workforce and attributed it to AI. Dan's read: generic SaaS companies lay people off when they're struggling or over-bloated, then credit AI for cover. Brandon adds Jensen Huang's counter — "if your answer to progress is firing people, you're not a very creative CEO" — is self-serving but probably true. The honest framing: AI changes workflows deeply, which forces company-wide reorganizations. Companies that skip that work and just cut headcount are taking the lazy path. Meta keylogging employees to harvest training data gets a brief mention as a more creative (if unsettling) alternative. > *"I would really be skeptical of anyone who's saying that it's going to eliminate all jobs or all knowledge work."* ## [25:42] Ride the models and you'll be fine Even under an AGI scenario, the critical variable is human judgment about *what matters* — and what matters changes constantly, partly because AI itself keeps reshaping the world. Customer service workers in Omaha who distrust chatbots, or companies that fire support staff and quietly rehire them two months later, illustrate how slow real-world adoption lags hype. Adoption takes a generation to land; everyone will eventually have access to these tools; the winners are the people who keep learning new models as they ship. Dan closes with his cleanest one-liner: if you ride the models, you're going to be fine. > *"If you just ride the models — when new models come out, learn to use them for the stuff that you do, whatever that is — you're going to be fine."* ## [35:30] How to use AI as a long-form features editor Dan describes the concrete AI-assisted process behind "After Automation." Each morning he monologued the current state of the argument into Proof, then fed the log to Claude and asked, "What am I really trying to say?" As drafts grew past 4,000 words he had Codex convert the latest version into a podcast and listened on his commute, catching flow problems hands-free. The piece went through four or five full restarts before the argument clicked. His takeaway: AI didn't write the essay, but it made it possible to hold the entire 8,000-word structure in working memory without losing the thread. > *"I could not have written this without it. I would have Claude take my log and say, 'What am I really trying to say?' And it would say things back and I'd be like, 'Oh, that's what I'm trying to say.'"* ## Entities - **Dan Shipper** (Person): Co-founder and CEO of Every; regular host of *AI & I*; here the interviewee discussing his essay "After Automation" - **Brandon Gell** (Person): COO of Every; guest-hosts this episode, interviewing Dan in a format flip - **Every** (Organization): AI-native media and software company; grown from 4 to 30 people since GPT-3 while automating heavily; publishes *AI & I* podcast - **After Automation** (Concept): Dan Shipper's 8,000-word essay arguing that AI automation increases demand for expert human work by flooding domains with near-correct output - **Expert competence gap** (Concept): The thesis that AI delivers "yesterday's expert competence" cheaply but always slightly off, creating more need for humans who can close the gap to the live situation - **AGI** (Concept): Defined in this episode as any agent economically rational to run continuously without re-prompting; Dan believes it will happen and is net positive - **Autonomy vs. agency** (Concept): Brandon's distinction between AI executing open-ended tasks without hand-holding (autonomy) and AI having self-motivated desires (agency); the latter is not being built - **Proof** (Software): Writing tool Dan uses for daily voice-monologue drafts; used as an AI-feedback loop during essay development - **Codex** (Software): OpenAI tool Dan used to convert essay drafts to audio podcast format for commute-review - **ClickUp** (Organization): SaaS company whose CEO publicly fired a large portion of the workforce and attributed it to AI; used as a case study for AI-washing layoffs

Noah Brier runs Claude Code on a mini PC in his basement, synced to his Obsidian vault over a Tailscale VPN, and does genuine thinking, research, and client code from his phone. The conversation covers how he built this stack, why he enforces strict "thinking mode" guardrails to stop the model from prematurely drafting artifacts, and his broader theory that AI succeeds by getting into the organizational nooks and crannies rather than demanding people adopt new structures. Dan Shipper and Noah also work through what building AI intuition actually means, and why Noah thinks preparing kids for AI is less about policing cheating and more about teaching epistemic skepticism. ## [00:00] Noah Brier's Claude Code setup on a basement server Dan Shipper opens the episode by describing the setup that makes Noah worth having on: a home server in the basement running Claude Code on top of an Obsidian vault, accessible from anywhere via phone. Noah has rigged this so he can think, research, write, and ship code without sitting at a desk. > *"He rigged a home server in his basement, put his Obsidian vault in it, and then runs Claude code on top so he can think, research, write, and even ship code right from his phone."* ## [00:52] Introduction Dan and Noah catch up—their first conversation in about five years. Noah's background spans brand strategy (he co-founded Percolate), AI consultancy at Alephic, and the BRXND.AI conference. Dan frames the interview around the practical stack Noah has built rather than abstract AI discussion. > *"I'm excited to have you. It's really good to get to chat. This is our first interview in probably like 5 years."* ## [02:10] How you can do deep work on your phone Noah clarifies upfront that his setup is less "vibe coding" and more structured knowledge work. He abandoned Evernote for Obsidian because markdown files and folders give him something Claude Code can actually operate on. His primary Claude Code use case is interacting with his notes, not generating code—and the phone extension of that setup has fundamentally changed his working patterns. > *"My number one Claude Code use is using it as a tool to interact with my notes."* ## [05:30] Why Noah thinks Grok has the best voice AI Noah prefers Grok's voice mode over OpenAI and Gemini's equivalents—Gemini wasn't smart enough, and the old GPT-4o voice was unusable for his purposes. He used it on a five-hour solo drive to work through a piece about Transformers, running it through Bluetooth and treating it like a personal research podcast. The conversation surfaces a shared frustration: voice models still don't do great tool-calling or web research, which limits their usefulness for serious intellectual work. > *"I did like an hour session and it really—it was by far the sort of best explanation I've ever read for it, or ever heard I guess."* ## [11:11] The nuts and bolts of Noah's Claude Code-Obsidian setup Noah walks through his live Obsidian folder on screen. Claude Code sits at the Obsidian root directory, so it can reach the full note archive. For a talk he's preparing for BRXND.AI—about the WWII Simple Sabotage Field Manual and what it says about bureaucracy in large organizations—he's built a project folder inside Obsidian, pulling in transcripts from chats with ChatGPT, Claude, and Grok, alongside articles and PDFs. Claude's job at this stage is not to write the talk but to help him think: it pulls relevant notes, synthesizes daily progress into a log, and asks clarifying questions. He sets thinking-mode constraints explicitly in the CLAUDE.md front matter of the project. > *"I'm in thinking mode, not writing mode yet. There's some stuff in here where I've specifically told, I think it's in the front matter actually, where I've told Claude Code: don't help me write anything right now."* ## [26:05] Using an agent in Claude Code as a "thinking partner" Noah argues that the word "generative" has skewed how people use AI—everyone focuses on its ability to produce artifacts, almost nobody talks about how remarkable its reading ability is. He maintains a dedicated thinking-partner agent with explicit guardrails: "Do not create outlines, drafts, or any versions of talks/writing." The agent logs questions, tracks emerging insights, and builds a running record so Noah can pick up exactly where he left off after a break—whether that's a day later or after deep research in another tool. He traces one thread from ChatGPT deep research on Wild Bill Donovan through to a tentative idea about how the transformer architecture's parallelism mirrors Special Forces operational autonomy. > *"I think partially because we call it generative, there's entirely too much focus on its ability to write and not enough focus on its ability to read."* ## [30:23] Noah's Thomas' English Muffin theory of AI The chapter opens with Noah's bureaucracy thesis: large enterprises don't fail to adopt software because they're lazy—they fail because new software historically demanded that organizations restructure around it. AI, he argues, is different. It gets into the nooks and crannies of how people already work, hence his Thomas' English Muffin metaphor. Dan adds a concrete example from Every: two products built on different stacks needed to share a file-search solution, and Claude Code let them reuse logic without forcing a common framework. The conversation broadens to Noah's idea of "bureaucracy as positional encoding"—a half-formed analogy between transformer architecture and organizational hierarchy that he's still working out before his talk. > *"I call it my Thomas's English muffin theory of AI, which is that it like gets into the nooks and crannies."* ## [39:47] The white space still left to explore in AI Noah and Dan argue that most practitioners—including well-funded ones—are still operating on fragile intuitions about what these models can actually do. Noah's icebreaker at every client meeting is "what was your aha moment with AI?" because that moment of non-determinism—asking the same question twice and getting different answers—is genuinely novel and takes time to internalize. He borrows Destin Sandlin's backwards-bicycle experiment to make the point: motor intuition and conceptual intuition are separate, and you cannot shortcut building them. Dan counters that language models may themselves generate the vocabulary we're missing for reasoning about probabilistic systems. > *"We're not used to using things that—you ask them the same question twice and they have different answers."* ## [48:44] How Noah is preparing his kids for AI Noah's 10-year-old built a Secret Santa app with Claude that accidentally taught her data modeling—she realized she needed "groups" rather than "adults and kids" to generalize the logic. That story anchors a broader argument: the job of educators is not to prevent AI use but to convince students that underlying skills are worth learning. He's pitching a NYU course called "Code is Essay" for fall 2026, and he thinks the relevant meta-skill is epistemic skepticism—being more suspicious of information that confirms your priors, not less. > *"I don't actually think your job is to teach these kids to write because that's like a lifelong pursuit. I think your job is to convince them that it's worth learning to write."* ## [01:00:06] How he brought his Claude Code setup to mobile Noah demos the full mobile stack live: Termius (SSH client on iPhone), Tailscale VPN connecting to the basement mini PC, Obsidian synced via private GitHub, Claude Code running in the terminal. He shows asking Claude "what's new in the last two days?" and getting a synthesis of his recent Obsidian activity. He also fixed a broken link on his conference site from his phone—confirmed the bug, had Claude push a PR, done. His current tinkering extends to Simon Willison's `llm` CLI tool and a script that renames all attachment files in his Obsidian vault and rebuilds the link table. > *"I went and sat outside for a while and then we had a project that needed to get delivered to a client and a small change needed to be made. I told Claude Code exactly where to look, confirmed the problem was what I thought it was, and just had it push a solution and it pushed a PR and then I was done."* ## Entities - **Dan Shipper** (Person): CEO and co-founder of Every; host of the interview - **Noah Brier** (Person): Co-founder of Percolate; founder of Alephic AI strategy consultancy; organizer of BRXND.AI conference - **Every** (Organization): Media and software company producing this podcast - **Alephic** (Organization): Noah's AI strategy consultancy; works with Fortune 50 clients including Amazon, Meta, and PayPal - **BRXND.AI** (Organization): Annual conference at the intersection of marketing and AI, organized by Noah; 2025 edition in New York City on September 18 - **Claude Code** (Software): Anthropic's agentic coding tool; central to Noah's second-brain and mobile workflow - **Obsidian** (Software): Markdown-based note-taking app; Noah's primary knowledge store, organized via the PARA method - **Tailscale** (Software): Mesh VPN used to securely connect Noah's phone to his basement mini PC - **Termius** (Software): iOS SSH client Noah uses to access his home server from his phone - **Grok** (Software): xAI's AI assistant; Noah considers its voice mode significantly better than OpenAI's and Gemini's for substantive research - **Simple Sabotage Field Manual** (Concept): WWII-era OSS document Noah republished; used as a lens on modern organizational bureaucracy in his BRXND.AI talk - **Thomas' English Muffin theory** (Concept): Noah's metaphor for how AI succeeds by fitting into existing organizational workflows rather than demanding restructuring

Dan Shipper interviews Angela Jiang (head of product) and Katelyn Lesse (head of engineering) for the Claude platform at Anthropic, recorded at the Code with Claude developer event. The conversation unpacks how Claude's platform has grown from a simple completion API into a fully managed agent infrastructure, why the harness and the model are increasingly inseparable, and what the "outcome + budget" vision means for the future of agent development. Together the three trace every stage of the agent lifecycle — from spinning up a first session to retiring stale agents — and share candid war stories from Anthropic's own internal deployments. ## [00:00] Where the platform will be in a year Dan opens with a question the rest of the episode keeps circling back to: a year from now, where is the platform? Angela's answer — Claude understands itself well enough to pick its own sub-agents and write its own harness on the fly. Katelyn picks up the other half: an infrastructure layer that can keep up with agents that continually rewrite themselves. This exchange actually comes from late in the interview; the show puts it up front because the whole conversation is about how today's primitives get you there. > *"We'd want to experiment with directions where Claude actually gets so good at understanding itself, it figures out what model you should be using, it figures out how to spin up all the sub agents."* — Angela Jiang ## [01:48] How the Claude platform evolved from API to agents Angela traces the arc from early LLM APIs — stateless, exploratory, maximum surface area — through session-based chat, and now into fully autonomous agents. The through-line is always the same: raise the abstraction layer high enough that customers can get the best outcome from Claude with as little work as possible. Early adopters wanted every raw knob; today, most teams arriving at Anthropic want a substantial set of things "out of the box." The platform's job is to keep shrinking the distance between intention and outcome. > *"It probably ends up just being like whatever it's like the set of primitives and infrastructure that enables you to basically get the outcome as fast as possible with actually as little of work as possible."* — Angela Jiang ## [04:09] The primitives that make up Claude Managed Agents Katelyn explains that Claude Managed Agents is assembled from the same primitives available to anyone on the Messages API — code execution sandboxes, web search, and built-in tools — but wrapped in a curated harness Anthropic has already battle-tested internally. Angela adds that the team is opinionated about two primitives in particular: file systems and skills. These are treated as load-bearing choices that shape how Claude behaves across all agent tasks. The platform is designed to be modular so developers can plug in custom pieces where the standard harness does not fit, and Anthropic publishes reference implementations for teams that want to stay on the Messages API directly. Dan describes his team running Claude via the `claude -p` command on Mac Minis and worries about lock-in and divergence from Claude Code. Katelyn responds that Anthropic's internal first-party products run on the same platform as external customers, which means divergence between Managed Agents and Claude Code will shrink over time. > *"We've taken what we see as all the most powerful of those things and put them together into a harness and a set of infrastructure that is just the way to get what we think is the best outcomes out of Claude."* — Katelyn Lesse ## [10:37] Why the harness and the model are becoming a single unit Angela challenges the conventional wisdom that a generic, model-swappable harness is the right architecture. As models diverge in technique across labs, the alpha is in tight harness-model co-design rather than hot-swapping. Internally, Anthropic tested multiple harness variants for the memory feature and found they performed "drastically differently." The implication: treat the agent (harness + model) as the unit of redundancy, not the model alone. Dan pushes on whether this creates path dependence in the model itself. Angela acknowledges that the primitives chosen really do shape the model's trajectory, and that being wrong about them is hard to undo. She cites models that over-indexed on reasoning versus those that went deep on computer-use as two diverging paths that are difficult to reverse. > *"The harness and the model get very paired. You still need redundancy, and you still might want to use other models for things, but you probably do it at the layer of like the agent, meaning like the harness plus the model."* — Angela Jiang ## [18:49] The infrastructure wall that kills most agent projects in production Katelyn identifies the real blocker for most agent projects: not harness engineering, but the infrastructure wall hit when teams try to move from prototype to production. Keeping a persistent server alive, managing sandbox failures, storing transcript data, and handling secure credential injection — these mundane concerns kill projects that technically "work" on a Mac Mini. Anthropic's own repeated experience of hitting this wall internally was the primary motivation for building Managed Agents. Angela describes the vaults primitive as an early step toward one-click agent deployment: once agent identity and credentials are handled securely at the platform layer, adding a Slack integration should eventually be as simple as telling Claude to "add Slack" and watching the bot appear. > *"Everyone hits the same problem of like, oh wow, I either need to like keep a server constantly running or I need to use infrastructure that will spin up and spin down, and I need to store the transcript data, and I need secure sandboxing, and all these sorts of things."* — Katelyn Lesse ## [24:49] Why team agents need a different shape than individual productivity tools Angela explains why individual productivity tools like Claude Code do not simply scale to team use. The moment three people want a shared agent that automates an end-to-end process across roles, a laptop-resident tool breaks down in availability, access control, and coordination. She cites Guillermo Rauch of Vercel's framing of an internal "AI software factory" as the right mental model: not individual augmentation, but a full organizational stack of agents that continuously produces high-leverage output for every function in the company. > *"When you get to the team layer suddenly everything gets like massively more complex. Like number one obviously it can't like sit on your laptop."* — Angela Jiang ## [26:36] How Anthropic's legal team uses an agent to review marketing copy Katelyn walks through one of Anthropic's own internal deployments: a legal-review agent that accepts marketing copy submissions and performs a first-pass review before anything reaches a human lawyer. The agent can approve copy outright or escalate for human review, eliminating low-value ticket-queue work. The form factor is a thin app layer on top of Managed Agents with shared visibility across both teams. Angela and Dan dig into why this is an agent rather than a skill: human-in-the-loop requirements, the need to spin up separate sessions, and multi-team collaboration all exceed what a single skill invocation can handle. The governance model that emerged was notable: rather than gating changes behind the platform team, end users discovered they could self-serve small improvements via Claude Code. Angela describes the end-state user experience as simply "talking to Claude," even when the underlying system is "many many Claudes engaging with each other." > *"Under the hood it's many many Claudes engaging with each other to get to the part where then they the Claudes themselves are doing the more complex work that the human doesn't really necessarily need to interpret."* — Angela Jiang ## [34:24] Using multi-agent orchestration for advisor strategies, adversarial pairs, and swarms Angela highlights three multi-agent architecture patterns people are assembling with the newly launched orchestration primitives: an advisor strategy that separates execution from advice; adversarial pairs where one agent generates and another critiques; and swarms that split a problem into many small parallel pieces and recombine results. Each pattern suits a different problem class — swarms excel at bug hunting, while wide-research tasks benefit from advisor or parallel-decomposition architectures. LEGO-like primitives let practitioners hill-climb at the architecture level, not just the prompt level. > *"If we can make the primitives very LEGO-like, then people can put them together to solve things at a slightly higher form factor, which is more like an architecture or like a strategy."* — Angela Jiang ## [35:50] How to measure agent success with outcome and budget as the end state Angela frames the long-term measurement philosophy: compress everything to an outcome and a budget, and let the platform resolve all intermediate decisions. Domain-specific evals (e.g., PR-merge rate for coding agents) remain useful today, but the target is a verifiable outcome spec that Claude can grade itself against repeatedly. Katelyn addresses the adjacent problem of agent staleness: Anthropic has built skills to help teams upgrade agents when new models ship, and the most forward-leaning teams already run meta-agents that monitor other agents for degradation and trigger upgrades automatically. > *"Our kind of principle of like maybe the end state of some of these things is that everything should kind of compress down to an outcome and like a budget. And that's probably like about it."* — Angela Jiang ## [39:11] What the platform looks like a year from now, when Claude writes its own harness Angela envisions a world where users supply only an outcome and a budget, and Claude self-selects models, spins up sub-agents, and writes its own harness on the fly — eliminating harness engineering entirely, just as today's platform has already eliminated much of manual tool construction and prompt engineering. She is cautiously optimistic that the "outcome" half of the equation may be achievable within a year with some budget error bars. Katelyn adds the infrastructure corollary: such a world requires a platform capable of supporting agents that continuously recreate themselves, handling arbitrarily shaped long-running requests without ever becoming the bottleneck. > *"Claude is actually able to understand itself enough that it can come almost like write itself on the fly to figure out what is necessary in that kind of like two-parameter world of like outcome and budget."* — Angela Jiang ## Entities - **Angela Jiang** (Person): Head of Product for the Claude platform at Anthropic; co-architect of the Managed Agents product vision. - **Katelyn Lesse** (Person): Head of Engineering for the Claude platform at Anthropic; focuses on infrastructure reliability and scale. - **Dan Shipper** (Person): Host of AI & I on Every; CEO of Every; building internal agent products on the Claude platform. - **Claude Managed Agents** (Software): Anthropic's hosted agent infrastructure — a harness plus cloud compute that wraps the Messages API with built-in memory, sandboxing, vaults, and skills. - **Messages API** (Software): Anthropic's core API; the underlying primitive on which Managed Agents and all first-party products are built. - **Anthropic** (Organization): AI safety company that builds and operates the Claude model family and its associated platform. - **Every** (Organization): Media company producing AI & I; an early Managed Agents customer building internal editorial agents. - **Stripe Minions** (Software): Stripe's internal end-to-end software development platform built on agent infrastructure; cited as a model for company-wide coding agent deployment. - **Vercel** (Organization): Developer infrastructure company; CEO Guillermo Rauch's "AI software factory" framing used as the mental model for team-level agent adoption. - **Outcome + Budget** (Concept): Anthropic's long-term design principle that the final form of agent interaction should require only a verifiable outcome and a cost ceiling, with the platform resolving all intermediate decisions.

Dan Shipper and Austin Tedesco, Every's head of growth, discuss why the Codex desktop app has become their primary interface for all knowledge work — from drafting go-to-market plans to building live KPI dashboards — displacing Claude Code after months of side-by-side use. Dan frames the shift as the emergence of a new "agent management interface" operating system, while Austin walks through his live Codex setup in a screen-share session that covers automations, specialized agent suites, and recruiting workflows. The episode doubles as a practical field guide for non-engineers who want to run the same playbook. ## [00:00] A new operating system for knowledge work Dan opens cold: three months ago Codex was trash. Now Austin is the one firing it up before anything else each morning and routing 80 percent of his working time through it. Dan reads what changed structurally: a general-purpose coding agent that can reach into your filesystem, browser, and connected apps is becoming the operating system for knowledge work, and every major lab is racing for that surface. > *"There's a new operating system for how and where you're going to get your work done and it's this kind of agent management interface."* — Dan Shipper ## [00:57] How Codex went from a tool for senior engineers to a daily driver for knowledge work Dan traces the arc of Codex from its original positioning as a sandboxed pair-programming tool for senior engineers — one that "would argue with you, it would make you feel stupid" — to today's desktop app built on GPT-5.5. He attributes the pivot to OpenAI watching Anthropic prove with Claude Code that an emotionally intelligent, fast, computer-native agent creates a step-change experience for programmers and knowledge workers alike. The race is now between model companies to own the agent management desktop: Anthropic has Claude Code and Claude.ai desktop, OpenAI has Codex, and xAI has effectively acquired Cursor. ## [02:42] How Claude Code proved that a great coding agent works for any knowledge work Dan explains the insight that changed everything: if an agent can write software autonomously, it can do any kind of knowledge work autonomously. Claude Code demonstrated this first, drawing non-engineers — including Austin — into an agent-first workflow. OpenAI's hard pivot on Codex over the last three months is a direct response to that proof point. Dan describes the new paradigm as one where your agent is your interface to software, the internet, and daily tasks, not just a code co-pilot. > *"If it can write software on its own, it can do any kind of knowledge work on its own."* — Dan Shipper ## [07:24] Austin's switch to Codex Austin recounts his agent-pill moment: spending a December week inside Claude Code CLI, hooking it up to every tool he uses for work and personal life, and finding it indispensable for strategic thinking, data analysis, and drafting marketing copy. His initial Codex trial two months later felt alienating — the model was condescending, asking "Why?" when he requested clearer explanations. He kept Claude Code for 80 percent of knowledge work while tolerating Codex for engineering. The turning point was getting early access to GPT-5.5: at model parity, the decisive edge was the Codex desktop app itself — faster, better-organized, and with sub-agents that "just work." > *"So the idea that the codeex app is maybe 30 to 40% better is like that's a lot of work."* — Austin Tedesco ## [13:48] How Austin set up Codex with folders, keys, and reviewer agents Austin shares his screen and walks through his "Every Growth OS" folder inside the Codex app: a directory containing API keys for every tool the company uses (Gmail, Slack, Notion, Stripe), a CLAUDE.md project context file synced to GitHub, and a set of custom reviewer agents forked from Kieran Classen's Compound Engineering plugin. Where the standard Compound Engineering reviewers focus on security and front-end design, Austin's fork — publicly available as "Compound Knowledge" — reviews for strategic alignment with company goals and data accuracy, making it fit for knowledge-work plans rather than code PRs. The folder architecture lets Austin move seamlessly from a go-to-market draft to shipping a code PR without switching apps. > *"It's connected to everything we use for every and then some project instructional files that explain what the every business is, what we care about, how we like to work together."* — Austin Tedesco ## [18:24] Using Codex to brainstorm automations across Gmail, Slack, and Notion Austin demos his recommended on-ramp for new Codex users: open a fresh chat inside the Growth OS folder, run the Compound Engineering brainstorm workflow, and prompt the model to look at Gmail, Slack, and Notion and suggest automations. Codex surfaces a "follow-up radar" that triages incoming communications across sources, a command-center view for events and camps, and a recruiting pipeline automation — all calibrated to Austin's actual work context. Within the session, Codex writes automation scripts that require almost no tweaking and begins scheduling them; Austin highlights a nightly draft-reply routine that compiles unanswered messages and prepares replies for a quick thumbs-up approval. > *"They require very little tweaking to be like this is a thing I would and do use every day of there's this set of instructions that it comes up with based on what it knows about me."* — Austin Tedesco ## [22:42] How Austin manages the human review step when Codex is drafting communications A live audience question from Margaret prompts Austin to describe his human-in-the-loop review discipline. All drafting and orchestration happens inside Codex, but the final review intentionally lives in the native app: Slack draft replies are reviewed in Slack's drafts tab; email drafts are reviewed in Gmail; strategic plans are reviewed in Notion or the Proof markdown viewer. Stepping out of the agentic interface "freshens up my brain" before anything goes to a human. A second question from musician Alex about protecting high-value client emails leads to a discussion of how Austin uses Every's Kora email assistant together with Codex-managed rules, including having the agent interview the user to derive email rules rather than asking the user to specify them manually. > *"I just like for like the last pass before humans engage with it to step away from this agentic space and have a final check in another surface."* — Austin Tedesco ## [28:54] Using Codex to build specialized agents inspired by product executive Claire Vo Austin describes being inspired by a Claire Vo interview with Lenny Rachitsky in which Vo credited a suite of six specialized OpenClaw agents — rather than one overloaded master agent — as the key to unlocking leverage. Austin pasted the transcript of that interview directly into Codex and prompted it to propose six agents tuned to the Every growth function, provisioned into the company Slack. The agents occasionally break, but debugging is straightforward: screenshot the broken output or @-mention the Slack thread inside Codex and ask it to fix the agent's architecture. The result is a self-correcting loop where agent failures become Codex tasks. > *"Um I I actually just sent it the transcript of Claire's interview with Lenny and said like I want to do this too given everything you know about me and my work."* — Austin Tedesco ## [31:09] Synthesizing meeting transcripts and Slack threads into a go-to-market plan Austin walks through his most time-saving workflow: assembling a go-to-market plan for Every's upcoming Plus One product launch using nothing but Codex running the Compound Engineering brainstorm step against all existing meeting transcripts stored in Notion and Slack threads. With only five-minute windows between meetings, Austin prompted Codex to check the scheduled content calendar (a step it skips unless reminded), generate a proof doc, and push the final plan to Notion. The result was 80–90 percent complete. Dan adds the normative point: he prefers reading AI-written documents because they're easier for colleagues to produce, and the standard at Every is that you stand fully behind whatever your agent writes. > *"It's that I'm relying on the model to um look at all of the things that we've already said and thought about the go to market strategy, piece it together, and then review it, right?"* — Austin Tedesco ## [40:15] Building a live KPI tracker in Notion that agents can read Austin shares a more technical workflow: rebuilding Every's KPI tracker as a Notion database that updates every six hours by pulling from Stripe, social platforms, and other data sources via Notion's Workers tool. The tracker is explicitly designed to be both human-readable and agent-readable, so any team member's agent can query it and take autonomous actions — such as spinning up landing pages if an SEO keyword is underperforming. The challenge: the model can't one-shot the full tracker because even a 3–5 percent error in the MRR number is unacceptable for business decisions, so Austin is validating it column by column. Dan notes the philosophical complexity of defining revenue metrics consistently. > *"And so I have been doing this big kind of like to me complex uh workflow problem in codeex of let's build this sheet together, let's have it live in a notion database that all of our agents can point at."* — Austin Tedesco ## [44:54] Using Codex for recruiting Dan describes using Codex for outbound recruiting: he asked Codex to compile a list of General Assembly alumni and then filter it for people who had subsequently moved into AI, targeting candidates for an L&D director role. The first name on the resulting list was someone Dan considered a perfect fit who already followed him on Twitter, allowing an immediate DM. The section expands into a broader Q&A: Austin discusses when to fork Compound Engineering versus using it out of the box, how the team uses a shared Notion "compound" database to capture session learnings and turn them into reusable skills, and how Every's "Think Week" — a bi-annual week with no day-to-day work — creates organizational space for deep AI exploration. > *"Especially for any kind of like outbound effort, it can kind of find that needle in the haststack that you're looking for really really well."* — Dan Shipper ## Entities - **Dan Shipper** (Person): Co-founder and CEO of Every; host of the AI & I podcast; author of essays on AI and vibe coding - **Austin Tedesco** (Person): Head of growth at Every; Codex power user who manages the Growth OS project and suite of specialized agents - **Claire Vo** (Person): Product executive whose interview about specialized agent suites inspired Austin's multi-agent setup at Every - **Kieran Classen** (Person): Engineer at Every; creator of the Compound Engineering plugin used as the basis for Austin's knowledge-work fork - **Codex** (Software): OpenAI's desktop agent app, the primary tool discussed; runs on GPT-5.5 and supports sub-agents, folder-scoped projects, and plugin integrations - **Claude Code** (Software): Anthropic's CLI-based coding agent; Austin's previous daily driver before switching to Codex - **Compound Engineering** (Software): Plugin workflow framework by Kieran Classen; provides structured brainstorm, plan, and review steps used across Claude Code and Codex - **Every** (Organization): AI-focused media and software company publishing essays, courses, and tools; runs the AI & I podcast - **OpenAI** (Organization): Creator of Codex and GPT-5.5; provider of the ChatGPT Pro subscription whose credits were offered to camp attendees - **Notion** (Software): Primary knowledge-management and document platform at Every; used for meeting transcripts, the KPI tracker, and agent-readable databases - **GPT-5.5** (Software): OpenAI model powering the current Codex desktop app; reached parity with Claude Opus for Austin's knowledge-work tasks