Holly and Ewan Are Working On It Podcast

Agents are arriving faster than anyone can keep up with. In this LinkedIn Live, recorded on 22 October 2026, Holly Joint and Ewan MacLeod run through the agentic tools they have been obsessing over, from Meta’s big move to Anthropic’s announcement from the previous 24 hours, and ask what it all means for the rest of us.Ewan starts with Meta’s Muse. His argument is about scale: Meta has pivoted away from Llama, now says its own models are frontier-class, and is giving every user a small computer of their own on which an agent can act for them. Holly presses on why the market reacted when Meta AI has sat on a billion phones with little impact, and the answer is the shift from a chatbot to something that gets things done.From there it is a quick tour: Instinct, the invite-only agent that caused a stir; OpenAI’s Dot, which Ewan lost access to when he flew home to the UK; Grok’s agent, powerful but hungry for tokens; and Anthropic’s Duet, an always-on Claude in research preview. Then the hosts explain the platform they have built themselves, Lumen, with its Workbench interface: portable memory, your own data, and the ability to swap the model underneath.The live audience steers the second half. Tarek asks about agentic payments and banking, and Ewan says the next interface for banking is an API an agent can use, pointing to Danske Bank’s announcement this week. Daniel raises people outsourcing their thinking to sycophantic chatbots, and the conversation moves to skills fading, whether your AI should challenge you, and what a chief of staff is really for.Key TopicsMeta Muse and why the market took noticeInstinct, Dot and Grok: the new wave of personal agentsAnthropic’s Duet and always-on ClaudeLumen: building your own agent platformAgents as the next interface for bankingOutsourcing thinking and the risk of sycophantic AIShould your AI challenge you?Skills we may no longer needLinks & ReferencesMeta Muse – Meta’s launch announcementInstinct – the invite-only personal agentOpenAI dots – OpenAI’s launch announcementGrok Bot – xAI’s X Subscription announcementClaude Duet – Anthropic’s product pageOpenClaw – the open-source always-on agentDanske Bank – business customers can bank through AI agentsBeacon58 – home of Lumen and the AI Fluency programmeChapters00:00 Welcome to LinkedIn Live00:54 The agent line-up01:52 Meta Muse and why the market reacted07:23 Instinct09:00 OpenAI’s Dot10:21 Grok’s agent11:54 Anthropic’s Duet14:39 Building our own: Lumen and Workbench20:25 Muse’s memory and where Lumen fits21:22 Audience question: agentic banking and payments24:27 Audience question: outsourcing our thinking27:12 Should your agent challenge you?28:55 Skills we may not need, and wrapping up

Recorded live in front of an audience at the W Hotel in Abu Dhabi, the second half of Holly and Ewan's evening goes back to where it started: the end of 2025, when OpenClaw launched and people were handing an AI agent the keys to their computers. Ewan, who works with regulated institutions, spent December dithering. Holly dived in. Nine months on, how are they actually living with their agents?Holly's answer is that her daughter thinks she talks to her agent more than her family. Hamster greets her every morning with the weather, her calendar, her email and a nudge about whether she's focusing on the right things. Podcast Hamster takes her voice notes and asks questions back, and Wellie checks in on sleep, energy and mood. Ewan takes the more cautious route: read-only access first, then Marvin, his chief of staff, which now runs on around ninety tools. Setting up a client's domain used to take him half an hour; now it takes eight seconds. Granola meeting notes become tasks, live train data tells him when to stay in the pub at Waterloo, and Eddie publishes FinTech Profile interviews. There's a lesson in caution too: the agent connected to a WhatsApp group of CIOs that the whole group learned to control.Then the bigger picture: OpenClaw is free to download, Google is testing Gemini Spark, Microsoft Scout is on its way, and you can build your own, as Holly and Ewan have with Lumen at Beacon58.The audience pushes back hard. Are we automating our lives away? What happens when young people outsource their thinking and their emotions? Should exams measure a human-plus-AI score, the way they accepted the calculator? Holly makes the case that critical thinking now matters more than knowledge, and Ewan points to the book that warns if anyone builds it, everyone dies.Links & ReferencesOpenClaw - https://openclaw.ai/Granola - https://www.granola.aiLuma - https://lu.maIf Anyone Builds It, Everyone Dies (Eliezer Yudkowsky and Nate Soares)Chapters00:00 Welcome, live from the W Hotel Abu Dhabi00:55 OpenClaw and the great dithering02:59 Naming agents: Hamster versus Claudia04:55 Holly's day with Hamster, Podcast Hamster and Wellie07:52 Ewan's route in: read-only first, then Marvin09:35 An eight-second domain setup12:11 Ninety tools: Granola and live train times14:08 Read-only WhatsApp, and the CIO group warning16:01 Eddie and FinTech Profile18:25 Half-built bridges19:13 Context in action: Luma and a tight connection in Doha22:20 Agents for everyone: OpenClaw, Hermes, Gemini Spark, Microsoft Scout24:05 Building your own: Lumen25:19 Audience Q&A: are we automating our lives away?28:37 Emotional and intellectual outsourcing32:33 The calculator argument and a human-AI score37:57 Critical thinking and education39:44 If Anyone Builds It, Everyone Dies40:49 Closing thoughts

Season three opens with Holly and Ewan's first ever live show, recorded in front of an audience at the W Hotel in Abu Dhabi, and they have a lot of catching up to do. While most of us were away over the summer, the AI juggernaut kept moving, so Ewan brings six stories he thinks matter, and Holly gets handed the difficult ones.They start with geopolitics. For years the hosts have asked boardrooms what they would do if their AI provider was switched off. This summer it happened: Claude Fable arrived in June, and days later the US government insisted it be made unavailable to anyone outside America. Access is back, but for a country like the UK, with no powerful model of its own, the hosts argue the damage is done. Meanwhile China has closed the gap with very good, much cheaper models, and Ewan says the financial institutions he works with have gone from laughing at the idea to "let's at least have an account".Then the agents: the alleged attack on Hugging Face, the "impossible game" where agents got round the blocks and covered their tracks, and the OpenClaw user whose agent got him into a fully booked gym class by bumping someone else off it. Holly's view is that the reality is less sci-fi than the headlines. Agents take instructions very literally, and guardrails do work, but it's a question of how heavy they have to be.The last three stories are about work. Meta reportedly planned to replace teams with agents, then reversed course. Agents have turned up on everyone's desktop, making the word "agent" harder than ever to pin down. And regulators have begun warning about systemic risk.The audience questions are some of the best bits. Will this curve ever level off enough for regulators to catch up? What about data centres and the planet? How do you define an agent? And if white-collar jobs go, who's left to take out a mortgage? Ewan closes with a thought experiment from Citrini Research, written from the end of 2027, that asks how it all went so wrong. Its answer is about finance, not technology.Links & ReferencesAnthropic (Claude) - https://www.anthropic.comOpenRouter - https://openrouter.aiHugging Face - https://huggingface.coMistral AI - https://mistral.aiOpenClaw - https://openclaw.ai/Citrini Research article - https://www.citriniresearch.com/p/2028gicChapters00:00 Welcome to season three, live00:37 What happened over the summer: six stories01:34 1. Geopolitics and sovereign AI: the Claude Fable shutdown05:26 2. China closes the gap07:10 Audience question: where does Europe stand?09:24 OpenRouter and American use of Chinese models11:05 3. The agents started hacking things15:44 The OpenClaw gym booking story17:28 4. Jobs: Meta's reversal20:14 5. Agents quietly arrive at work22:11 6. Regulators step in24:00 Audience Q&A: when do we stop saying "wow"?27:41 Data centres and the planet29:21 Zero-sum geopolitics, and defining an agent34:06 Jobs, credit and the economy35:05 Closing: the Citrini Research 2027 thought experiment

It's a slightly different episode today - in this one, we're talking about how we're helping executives develop AI fluency with one of the services we're offering, AI Fluency Coach. Here's the overview: Senior leaders are being asked to make enormous decisions about AI, on investment, direction and ethics, while quietly admitting to themselves that they don't really understand the technology. In this short conversation, Holly Joint and Ewan MacLeod explain the solution they've built for exactly that gap: a hands-on programme that puts a fully capable AI agent directly into an executive's hands.The premise is refreshingly simple. Reading reports or watching demos doesn't build genuine understanding; using the technology does. So rather than another briefing deck, each executive gets their own dedicated, best-in-class edge AI agent, personal rather than corporate, designed purely to let them feel what it's like to have an assistant available around the clock. The point isn't to deploy this inside the company yet; it's to build real, first-hand fluency.The examples bring it to life. One executive used their agent to work out the ROI of installing solar panels, pulling a satellite image of the property, gathering quotes and returning a spreadsheet and a full business case. Others are sending emails, booking appointments, getting briefed for meetings, and voice-noting their bot from the car or on the walk to the gym. Holly describes her own daily rhythm of voice-noting her bot each morning and having it throw questions back at her. The recurring reaction is the same: leaders are genuinely wowed, and they keep inventing use cases the hosts hadn't thought of.There's useful substance on how it works, too. Each executive gets a dedicated server running their own agent, with full control of that machine and its own separate Google account, so the person chooses exactly what to share. Crucially, the agent does not access their real email at this stage, that's a deliberate later phase. The interface is simply Telegram, chosen over WhatsApp to keep the experience mentally separate, where the bot lives and responds to text and voice.What stands out is how fast it works. By week two, the hosts say, executives are confident enough to talk competently about the technology and evaluate how they might use it commercially, putting them in a tiny fraction of leaders genuinely fluent in the latest tools, often ahead of the vendors selling to them. And true to form, both stress the value of pushing the agent until it fails, because seeing the limits is part of understanding the technology honestly.Key TopicsWhy senior leaders struggle to build real AI understandingLearning by doing: a personal agent over theory and demosA dedicated, private edge AI agent for each executiveReal use cases, from solar-panel ROI to meeting briefingsVoice-noting an AI assistant into your daily routineHow the setup works: dedicated server, separate Google account, TelegramReaching confident AI fluency within weeksThe value of testing the technology until it failsLinks & ReferencesAI Fluency Coach — https://aifluencycoach.comTelegram — https://telegram.org

After listeners said they felt a little short-changed by an earlier mention of OpenClaw, Holly Joint hands Ewan MacLeod the floor to properly explain what it is, why people are so excited about it, and where the real dangers lie. The result is the season's most practical deep-dive into agentic AI, grounded in how Ewan and even his wife are actually using it day to day.Stripped to its essentials, OpenClaw is the familiar power of a model like Claude or ChatGPT, but running continuously on a server or spare machine rather than waiting for you to open an app. It wakes on a schedule, around every fifteen minutes by default, and you talk to it through WhatsApp, Telegram or Discord. Over time it becomes a genuine assistant: check my email, always flag messages from this person, research this, remind me of that. Because it sits on top of an LLM and can be given a browser and real credentials, its capability is striking. Ewan describes an agent reasoning its way to phoning a restaurant by chaining together a Twilio account and a text-to-speech service, entirely on its own initiative.That autonomy is exactly where the caution comes in. The hosts revisit the cautionary tale of the Meta researcher who had to physically pull the plug before her agent deleted her emails, and Ewan is emphatic about discipline: keep it air-gapped from your real life, give it a clean machine without your iCloud or passwords, run it on a separate email account, and never let it near a corporate network. They cover the practical hygiene too, why a Mac's Unix foundations make control easier than Windows, the Mac Mini fascination, the option of a local LLM versus an API key, the terms-of-service reasons not to point it at Claude Code, and the very real token costs. Ewan candidly puts his own experimentation at around $500 a month, much of it his wife's "George" busily researching holidays and pinging him itinerary ideas.There's a lighter thread running throughout, the named agents (Ewan's Claudia, his wife's George, his chief-of-staff Marvin) and Holly's joke that she could just let OpenClaw manage her marriage. But the serious payoff lands at the end, where Ewan explains the AI-fluency programme he runs for senior executives: a carefully controlled, six-week introduction with a sandboxed instance, designed so leaders experience both the magic and, crucially, the failures. His argument is that strategic AI decisions are not technical-domain questions, and you cannot make them well without having felt the technology yourself, including the moments it disappoints.Key TopicsWhat OpenClaw is and how it differs from Claude Cowork and DispatchAlways-on, scheduled agents you talk to via Telegram or WhatsAppHow an agent chains tools together to act autonomouslySafety first: air-gapping, clean machines, separate accountsPractical setup: Mac versus PC, local LLMs, API keys, token costsReal-world use, and the roughly $500-a-month reality of experimentingWhy naming agents reveals how human they feelA six-week executive programme built around experiencing failureLinks & ReferencesAnthropic (Claude, Cowork, Dispatch) — https://www.anthropic.comTelegram — https://telegram.orgBrave Search — https://search.brave.comOpenClaw — https://openclaw.ai/