Skip to content

Podcast

Ralph Wiggum, Clawdbot, and Mac Minis- How Pros Are Vibe Coding in 2026

The AI Daily Brief: Artificial Intelligence News and Analysis

Source ↗ ← All highlights
  • Build With Tools Instead Of Debating Them
    • Ignore spectacle and focus time on learning to build with AI tools to understand practical impacts.
    • Use hands-on projects instead of debating Davos-style conversations to prepare for change. Transcript: Nathaniel Whittemore Yet, of course, we live in the world that we live in, and like it or not, the conversations that happen in Davos are a useful reflection on what global leaders think about this moment, And so give us insight into the context in which this industry and this technology is going to operate. One side of the conversation was the voices coming from the tech industry. Reuters summed up that voice as jobs, jobs, jobs. The AI mantra in Davos as fears take a back seat. Now that is a specific reference to NVIDIA’s Jensen Huang, who basically made the argument that the amount of demand for chips, the infrastructure layer that needs to be built, the Energy infrastructure that needs to be built to service it, is all a big moment of job creation. And indeed, I think it is the case that fairly uniquely relative to other moments of creative destruction, even the transitional moment has the potential for a lot of creation as well. I think Jensen is right to identify that there is a lot more skilled labor outside of knowledge work that needs to be developed for this transition. In other places, tech leaders talked about the productivity benefits that they were seeing. Cisco talked about projects that had been too tedious to even contemplate before that could now be done in a couple of weeks. IBM’s chief commercial officer, Rob Thomas, said that AI was at the ROI stage. He told Reuters, you can truly start to automate tasks and business processes. TechCrunch said that even though we anticipated AI being a big topic of conversation, the extent to which it shaped the event, with even the physical surrounding being dominated by Tech companies and pavilions, was notable. And yet, of course, if the technology folks were excited, concerns about AI-related job displacement were on the agenda as well. Christy Hoffman, the general secretary of the 20 million member strong Uni Global Union, said AI is being sold as a productivity tool, which often means doing more with fewer workers. International Monetary Fund Managing Director Kristalina Georgieva called AI a tsunami hitting the labor market with the potential to transform or eliminate 60% of jobs in advanced Economies and 40% globally. Now, I remember a study from a couple of years ago from one of the big global institutions, IMF or World Bank or one of them, that basically had those numbers, so I assume that’s what she’s Talking about. Providing some bright spot, she thought that as high-skilled workers see their wages rise because of AI, they would likely consume more in ways that benefited the local service economy. She said one in 10 jobs is already enhanced by AI, and the people in these jobs are paid better. When they’re paid better, they spend more money in the local economy. They spend more money in restaurants here or there. Demand for low-skilled jobs goes up, and actually total employment seems to slightly increase because of it. Now, for those who might be skeptical of this or seem like it feels relatively Pollyannish, there have been studies that have shown that, for example, in San Francisco for each new local Tech job, 4.4 jobs for positions like retail clerks, cooks, teachers, and dentists is also created. At the same time, the IMF still has some big concerns. The two that stood out is stagnating middle-class wages, especially for jobs that are not enhanced by AI, and increasing barriers to youth employment as AI takes over the entry-level Tasks. Now, behind the scenes in Davos, there was also a lot of jockeying for position. The Information wrote a piece all about how some Davos meetings were part of what seems to be a larger strategy for OpenAI to get more aggressive about its enterprise recruitment. Now, this effort was not strictly restricted to Davos. In fact, last week in San Francisco, Sam Altman hosted an extended business dinner with Disney CEO Bob Iger and other corporate execs. The information writes that the gathering was intended to preview a new OpenAI offering aimed at large companies, but they could not determine what that offering was. All that was happening while OpenAI COO Brad Lightcap and new chief revenue officer Denise Dresser were schmoozing over in Davos. Clearly, the company is trying to message that they are, in fact, not behind when it comes to enterprise. In a Davos session, OpenAI CFO Sarah Fryer said that by the end of the year, approximately 50% of their business will come from enterprise customers. And Sam Altman tweeted that they had added more than a billion in ARR over the last month just from their API business. Very clearly trying to shift the narrative, he says, people mostly think of us as ChatGPT, but the API team is doing amazing work. So what does this all add up to? It’s kind of hard to tell. Part of the reason that we may not be able to have quite as strong a sense of what the general sentiment around AI was, is just that there were, of course, other more geopolitical conversations That made even the AI conversation take a back seat. I think, if anything, Jamie Dimon’s crisp realism that no one can put their head in the sand, that AI is not a force that is likely to be stopped, but that there could be challenges for how Fast it’s going to cause change in society that we may have to address, might be a fairly good representation of the median. (Time 0:02:25)
  • Davos: Jobs, Productivity, And Enterprise Push
    • Davos highlighted both job-creation from infrastructure demand and worries about displacement and wage shifts.
    • Enterprise adoption is ramping fast, with OpenAI pushing to make half its revenue from enterprise this year. Transcript: Nathaniel Whittemore Reuters summed up that voice as jobs, jobs, jobs. The AI mantra in Davos as fears take a back seat. Now that is a specific reference to NVIDIA’s Jensen Huang, who basically made the argument that the amount of demand for chips, the infrastructure layer that needs to be built, the Energy infrastructure that needs to be built to service it, is all a big moment of job creation. And indeed, I think it is the case that fairly uniquely relative to other moments of creative destruction, even the transitional moment has the potential for a lot of creation as well. I think Jensen is right to identify that there is a lot more skilled labor outside of knowledge work that needs to be developed for this transition. In other places, tech leaders talked about the productivity benefits that they were seeing. Cisco talked about projects that had been too tedious to even contemplate before that could now be done in a couple of weeks. IBM’s chief commercial officer, Rob Thomas, said that AI was at the ROI stage. He told Reuters, you can truly start to automate tasks and business processes. TechCrunch said that even though we anticipated AI being a big topic of conversation, the extent to which it shaped the event, with even the physical surrounding being dominated by Tech companies and pavilions, was notable. And yet, of course, if the technology folks were excited, concerns about AI-related job displacement were on the agenda as well. Christy Hoffman, the general secretary of the 20 million member strong Uni Global Union, said AI is being sold as a productivity tool, which often means doing more with fewer workers. International Monetary Fund Managing Director Kristalina Georgieva called AI a tsunami hitting the labor market with the potential to transform or eliminate 60% of jobs in advanced Economies and 40% globally. Now, I remember a study from a couple of years ago from one of the big global institutions, IMF or World Bank or one of them, that basically had those numbers, so I assume that’s what she’s Talking about. Providing some bright spot, she thought that as high-skilled workers see their wages rise because of AI, they would likely consume more in ways that benefited the local service economy. She said one in 10 jobs is already enhanced by AI, and the people in these jobs are paid better. When they’re paid better, they spend more money in the local economy. They spend more money in restaurants here or there. Demand for low-skilled jobs goes up, and actually total employment seems to slightly increase because of it. Now, for those who might be skeptical of this or seem like it feels relatively Pollyannish, there have been studies that have shown that, for example, in San Francisco for each new local Tech job, 4.4 jobs for positions like retail clerks, cooks, teachers, and dentists is also created. At the same time, the IMF still has some big concerns. The two that stood out is stagnating middle-class wages, especially for jobs that are not enhanced by AI, and increasing barriers to youth employment as AI takes over the entry-level Tasks. Now, behind the scenes in Davos, there was also a lot of jockeying for position. The Information wrote a piece all about how some Davos meetings were part of what seems to be a larger strategy for OpenAI to get more aggressive about its enterprise recruitment. Now, this effort was not strictly restricted to Davos. In fact, last week in San Francisco, Sam Altman hosted an extended business dinner with Disney CEO Bob Iger and other corporate execs. The information writes that the gathering was intended to preview a new OpenAI offering aimed at large companies, but they could not determine what that offering was. All that was happening while OpenAI COO Brad Lightcap and new chief revenue officer Denise Dresser were schmoozing over in Davos. Clearly, the company is trying to message that they are, in fact, not behind when it comes to enterprise. In a Davos session, OpenAI CFO Sarah Fryer said that by the end of the year, approximately 50% of their business will come from enterprise customers. And Sam Altman tweeted that they had added more than a billion in ARR over the last month just from their API business. (Time 0:02:43)
  • Autonomy Is The Real Shift In 2026
    • Agentic coding is advancing by extending autonomy so agents run with minimal human intervention.
    • Teams focus on removing humans as bottlenecks to let agents work while creators do other things or sleep. Transcript: Nathaniel Whittemore Welcome back to the AI Daily Brief. Today we are doing a little bit of a catch-up on the terms that you might have heard in passing, especially if you’ve been anywhere near AI Twitter slash X over the past couple of weeks. There are a few things that might sound like absolute Greek to you, but which combined tell the story of how vibe coding, which I really mean AI and agentic coding, are evolving early Into this year. Entrepreneur and creator Riley Brown recently tweeted, cool Claude stuff, re-motion skill, Claudebot, C-L agent SDK, Ralph, and cowork. Now, if you are thinking, I don’t know what any of those things mean, don’t worry, you are not alone, and we’re going to get into much of it today. The context of all of this is the big shift in perception over the last couple of weeks, which has been pretty well chronicled in episodes throughout this month. It wasn’t that we got a new model or anything like that. It’s that everyone went home for the holidays, had just a little bit of downtime to start playing around, started working on some personal or professional projects with Opus 4.5 or Claude Code or 5.2 Codex or some combination thereof, and realized that what we could do with agentic coding was much, much farther than they might have thought. This was reinforced a couple weeks later when Anthropic dropped Claude Cowork, which is sort of like Claude Code for the rest of us, and revealed that it had been written 100% by Claude Code in just about 10 days. Now, if you want even more of a primer, suggest one of my previous episodes, Why Everybody is Obsessed with Cloud Code, Cloud Cowork is Cloud Code for Everybody Else, or most recently And probably most importantly, Why Code AGI is Functional AGI and it’s here. So that’s the setup, and we just keep getting evidence of how much things have shifted. Cursor CEO Michael Truel posted about a week and a half ago, we built a browser with GPT 5.2 in Cursor. It ran uninterrupted for one week. It’s 3 million plus lines of code across thousands of files. The rendering engine is from scratch in Rust with HTML parsing, CSS cascade, layout, text shaping, paint, and a custom JSVM. It kind of works. It still has issues and is of course very far from WebKit and Chromium Parity, but we were astonished that simple websites render quickly and largely correctly. And to be clear, this was an experiment in autonomy. While at first blush people thought it was one agent writing 3 million lines of code, it wasn’t. It was actually hundreds of concurrent agents. Cursor wrote it up in a blog post called Scaling Long-Running Autonomous Coding, and it’s very clear that Cursor is interested in pushing this frontier. They wrote, we’ve been experimenting with running coding agents autonomously for weeks. Our goal is to understand how far we can push the frontier of agentic coding for projects that typically take human teams months to complete. And indeed, if you want to take a step back and just try to understand psychologically where the vanguard of AI and agent decoders are right now, it is really all about pushing the boundaries On autonomy, breaking out, in other words, of being the bottleneck where without your consistent prompting, the AI isn’t doing anything. The leading agent decoders are in the midst of trying to build systems that work all the time with extremely minimal input from them. They want nothing less than armies of agents that work while they sleep. And that army idea is operative. In that same Cursor blog, they write, today’s agents work well for focused tasks but are slow for complex projects. The natural next step is to run multiple agents in parallel, but figuring out how to coordinate them is challenging. Initially, Cursor gave their coding agents equal status, and as they put it, let them self-coordinate through a shared file. Each agent would check what others were doing, claim a task, and update its status. Ultimately, however, this failed. The locking mechanism they implemented to prevent two agents from grabbing the same task ended up becoming a bottleneck. As they put it, 20 agents would slow down to the effective throughput of two or three with most time spent waiting. They tried a second strategy, where agents could read state freely, but rights would fail if the state had changed since they last read it. In other words, they couldn’t make different updates to the same code at the same time in an attempt to avoid conflicts. However, Kurser wrote this didn’t work either. Quote, as they put it, with no hierarchy, agents became risk averse. They avoided difficult tasks and made small safe changes instead. No agent took responsibility for hard problems or end-to implementation. This led to work churning for long periods of time without progress. The next approach they took was to separate roles. Instead of a flat structure, they created a pipeline where a subset of agents called planners would continuously explore the codebase and create tasks, and workers would pick up those Tasks and focus entirely on completing them. The workers, they wrote, don’t coordinate with other workers or worry about the big picture. They just grind on their assigned task until it’s done, then push their changes. At the end of each cycle, a judge agent determined whether to continue, then the next iteration would start fresh. This, they said, solved most of our coordination problems and let us scale to very large projects without any single agent getting tunnel vision. Now, this is the point at which they instituted this ambitious goal of building a web browser from scratch. Now, as we heard at the beginning, this worked, but not without a lot of challenges. They write, our current system works, but we’re nowhere near optimal. Planners should wake up when their tasks complete to plan the next step. Agents occasionally run for far too long. We still need periodic fresh starts to combat drift and tunnel vision. (Time 0:09:40)
  • Browser Built By Hundreds Of Agents
    • Cursor ran hundreds of concurrent agents to build a browser with millions of lines of code over a week.
    • They used planners, workers, and judge agents to coordinate progress and scale without a single human steering every change. Transcript: Nathaniel Whittemore Cursor CEO Michael Truel posted about a week and a half ago, we built a browser with GPT 5.2 in Cursor. It ran uninterrupted for one week. It’s 3 million plus lines of code across thousands of files. The rendering engine is from scratch in Rust with HTML parsing, CSS cascade, layout, text shaping, paint, and a custom JSVM. It kind of works. It still has issues and is of course very far from WebKit and Chromium Parity, but we were astonished that simple websites render quickly and largely correctly. And to be clear, this was an experiment in autonomy. While at first blush people thought it was one agent writing 3 million lines of code, it wasn’t. It was actually hundreds of concurrent agents. Cursor wrote it up in a blog post called Scaling Long-Running Autonomous Coding, and it’s very clear that Cursor is interested in pushing this frontier. They wrote, we’ve been experimenting with running coding agents autonomously for weeks. Our goal is to understand how far we can push the frontier of agentic coding for projects that typically take human teams months to complete. And indeed, if you want to take a step back and just try to understand psychologically where the vanguard of AI and agent decoders are right now, it is really all about pushing the boundaries On autonomy, breaking out, in other words, of being the bottleneck where without your consistent prompting, the AI isn’t doing anything. The leading agent decoders are in the midst of trying to build systems that work all the time with extremely minimal input from them. They want nothing less than armies of agents that work while they sleep. And that army idea is operative. In that same Cursor blog, they write, today’s agents work well for focused tasks but are slow for complex projects. The natural next step is to run multiple agents in parallel, but figuring out how to coordinate them is challenging. Initially, Cursor gave their coding agents equal status, and as they put it, let them self-coordinate through a shared file. Each agent would check what others were doing, claim a task, and update its status. Ultimately, however, this failed. The locking mechanism they implemented to prevent two agents from grabbing the same task ended up becoming a bottleneck. As they put it, 20 agents would slow down to the effective throughput of two or three with most time spent waiting. They tried a second strategy, where agents could read state freely, but rights would fail if the state had changed since they last read it. In other words, they couldn’t make different updates to the same code at the same time in an attempt to avoid conflicts. However, Kurser wrote this didn’t work either. Quote, as they put it, with no hierarchy, agents became risk averse. They avoided difficult tasks and made small safe changes instead. No agent took responsibility for hard problems or end-to implementation. This led to work churning for long periods of time without progress. The next approach they took was to separate roles. Instead of a flat structure, they created a pipeline where a subset of agents called planners would continuously explore the codebase and create tasks, and workers would pick up those Tasks and focus entirely on completing them. The workers, they wrote, don’t coordinate with other workers or worry about the big picture. They just grind on their assigned task until it’s done, then push their changes. At the end of each cycle, a judge agent determined whether to continue, then the next iteration would start fresh. This, they said, solved most of our coordination problems and let us scale to very large projects without any single agent getting tunnel vision. Now, this is the point at which they instituted this ambitious goal of building a web browser from scratch. Now, as we heard at the beginning, this worked, but not without a lot of challenges. They write, our current system works, but we’re nowhere near optimal. Planners should wake up when their tasks complete to plan the next step. Agents occasionally run for far too long. We still need periodic fresh starts to combat drift and tunnel vision. But the core question, can we scale autonomous coding by throwing more agents at a problem, has a more optimistic answer than we expected. Hundreds of agents can work together on a single code base for weeks, making real progress on ambitious projects. (Time 0:11:22)
  • Hierarchy Beats Flat Coordination
    • Flat peer agents became risk-averse and stalled on hard problems, so hierarchy improved throughput.
    • A pipeline of planners, workers, and a judge let many agents scale without tunnel vision. Transcript: Nathaniel Whittemore In that same Cursor blog, they write, today’s agents work well for focused tasks but are slow for complex projects. The natural next step is to run multiple agents in parallel, but figuring out how to coordinate them is challenging. Initially, Cursor gave their coding agents equal status, and as they put it, let them self-coordinate through a shared file. Each agent would check what others were doing, claim a task, and update its status. Ultimately, however, this failed. The locking mechanism they implemented to prevent two agents from grabbing the same task ended up becoming a bottleneck. As they put it, 20 agents would slow down to the effective throughput of two or three with most time spent waiting. They tried a second strategy, where agents could read state freely, but rights would fail if the state had changed since they last read it. In other words, they couldn’t make different updates to the same code at the same time in an attempt to avoid conflicts. However, Kurser wrote this didn’t work either. Quote, as they put it, with no hierarchy, agents became risk averse. They avoided difficult tasks and made small safe changes instead. No agent took responsibility for hard problems or end-to implementation. This led to work churning for long periods of time without progress. The next approach they took was to separate roles. Instead of a flat structure, they created a pipeline where a subset of agents called planners would continuously explore the codebase and create tasks, and workers would pick up those Tasks and focus entirely on completing them. The workers, they wrote, don’t coordinate with other workers or worry about the big picture. They just grind on their assigned task until it’s done, then push their changes. At the end of each cycle, a judge agent determined whether to continue, then the next iteration would start fresh. (Time 0:12:56)
  • Use Ralph Loop: Break Work Into Atoms
    • Break projects into atomic user stories with clear acceptance criteria before looping agents.
    • Use short RALF iterations, log learnings, then human-test edge cases to finalize features. Transcript: Nathaniel Whittemore So the idea of Ralph as applied to AI coding was described by developer Ryan Carson in a post on X. He writes, everyone is raving about Ralph. What is it? Ralph is an autonomous AI coding loop that ships features while you sleep. Each iteration is a fresh context window. Memory persists via Git history and text files. Now he gets into exactly what this loop looks like from a technical perspective, but the Startup Ideas podcast with Greg Eisenberg had Ryan on to explain it even more simply. And here’s how they summed it up. Step one, write a detailed PRD. That’s a product requirements document, which is a document that defines the purpose, features, functionality, and behavior of any new project or feature. It’s going to define why the product is being built, what success looks like, detailed requirements of what it should do, things like that. Now, after you write that detailed PRD, you’re going to convert it to extremely small, discrete, atomic, to use their words, user stories. Step three is that for each of those atomic units, you add clear acceptance criteria. Step four is looping your AI agent through each story. In step five, it logs learning so it doesn’t repeat mistakes. Step six, the person who initiated the RALF loop wakes up, tests it, and fixes the edge cases. Basically, the idea is to break down a complex project into very discrete smaller units that the coding agent can take on one by one, testing and looping until it’s finished and moving On to the next. (Time 0:16:51)
  • Claudebot Brings Local, Self-Improving Agents
    • Claudebot runs locally on personal hardware and connects apps via a gateway to act autonomously.
    • Its self-improving skills let users ask for new capabilities which the bot can often implement itself. Transcript: Nathaniel Whittemore A post on starryhope.com reads, at its core, ClaudeBot is an open source AI agent that runs on your own hardware. Unlike ChatGPT or Claude’s web interfaces, which process everything on remote servers, Claudebot operates locally with a gateway that connects AI models to the apps and services You already use. It can talk to you through WhatsApp, Telegram, Discord, Slack, Signal, and even iMessage. But the real magic is what it can do once it’s running. Given the right permissions, Claudebot can browse the web, execute terminal commands, write and run scripts, manage your email, check your calendar, and interact with any software On your machine. Perhaps the most compelling feature is that CloudBot is self-improving. Tell it you want a new capability, and it can often write its own skill or plugin to make it happen. One user wanted access to university course assignments. He asked CloudBot to build a skill for it. CloudBot did and then started using it on its own. Now, some are a little skeptical. Former NVIDIA engineer Boyan Tungus said, I’m as excited as the next guy about the possibilities of CloudBot running on a cluster of small local mini computers. (Time 0:19:16)
  • Mac Mini As A 24/7 Digital Employee
    • Nat ran ClaudeBot on a Mac mini as an always-on digital employee handling tests, Sentry errors, and customer workflows.
    • He woke to reports of fixes and saw the bot autonomously build support and success processes. Transcript: Nathaniel Whittemore And Nat would know, because he went viral when he posted a picture of a Mac Mini about a week ago and said, hired my first employee today. He followed up writing, yeah, this was 1000% worth it. Separate Claude, that’s the C-L version, plus Claude, the C-L managing Claude code and codex sessions I can kick off anywhere, autonomously running tests on my app and capturing errors Through a Sentry webhook, then resolving them and opening PRs. Basically, Nat has this set up to be working around the clock on a new agent that he’s building to automate agency-level content creation. On Saturday morning, Nat posted, Nothing like waking up to a report from ClaudeBot about everything that went wrong in my app yesterday and what it already did to fix it. A couple hours later, Nat was still going. He wrote, Built a customer success and support workflow for ClaudeBot now too. Analyzes transcripts from the day, emails customers with bad experiences apologizing and asking for any other feedback, adds their feedback to the daily report for our next morning Brainstorm. Basically, he’s got a digital employee that lives in a Mac mini, uses Claude code Opus 4.5 and Codex 5.2, and which he communicates with via Telegram. (Time 0:20:31)