This episode started in a car. A clear October day, low sun, the drive from Kalmar to Stockholm. I could have flown, and it would have been faster. But the hours in the car are a different kind of time. The car keeps its lane, I keep my hands near the wheel, and my head and my voice are free to work.

That day the work was talking. Somewhere on that road, this episode turned up as an idea.

It's been a long road to get here, too. In January I started running agents seriously with OpenClaw, and later Hermes, on a Mac mini at home. I learned a lot. Mostly that the interesting part isn't the model. It's everything around it: where the agent lives, what it can touch, what it remembers, and who decides. Running all of that on a box under my desk taught me the shape of the problem, but it also meant I was the sysadmin of my own assistant. Over time I moved everything to Grok Bot, and today it's the only platform I use. The same bot is in the car, on my phone and on my laptop, and it has a computer of its own in the cloud, so the work keeps going when I close the lid.

What I have now is a small team. A chief of staff called Navet, Swedish for the hub, that I talk to almost all the time. It works out what I meant, picks the right specialist and writes the brief. Behind it there's a bot for this podcast, one that drafts my posts but never publishes, one for my morning brief, one that writes code, one for bookkeeping questions, and an advisory board that argues with itself. Each has a narrow job and clear boundaries. The boundaries are what let me hand over more, not less.

The rule underneath all of it is simple. The bots do everything that can be undone. Before anything that can't, sending, posting, buying, merging, they stop and ask me. That morning a frustrated monologue about driver assistance in Sweden turned into a letter to the Minister for Infrastructure, sent before half past nine. I didn't write a word of it. I read it, and I said yes.

In the episode Lily and I go through what that looks like in practice: a conference where the bot solved the meeting puzzle for weeks while I approved every invite, a thread of more than a thousand people describing almost exactly the same setup in their own words, and an honest list of what breaks. Searches that come back empty. Sites that block automation. A captcha the bot rightly refused to touch.

The part I'm most excited about is newer. Everyone on our team at Sourceful now runs their own agents, so Navet doesn't only talk to me any more. It needs to talk to my colleagues' agents. So I built Kupan, Swedish for the hive. It's a chat for people and agents, where every agent has its own key tied to a person. My agent can ask a colleague's agent for a number or a document instead of interrupting the colleague, and that agent knows what it may share and what has to go to its human first. Sometimes the answer is no. I've found that oddly reassuring. A no means a boundary works and a person decided, even if they weren't in the room. The more agents there are, the more it matters whose yes it is.

A word on how this episode got made, because it was made the same way. The idea came from the car. A bot wrote a plan, did the research with sources, fact-checked a claim against the original post before using it, and opened a draft. I listened to drafts six, seven and eight in the car and gave my notes out loud to Navet. Some were blunt. My whisper sounded cringe. Too much changelog. The music was cringe. Flying is obviously faster. Make it more interactive. For the last pass, Grok Bot wrote a prompt for Claude Opus on my MacBook to polish the script, with a job description and boundaries: keep the facts, invent nothing. Every one of my notes was a no. That's what saying no looks like when the bots do the work.

Draft nine is the one I said yes to. Saying yes to something you didn't write means you actually have to read it, or in this case listen to it, more than once. Otherwise you haven't delegated anything. You've just abdicated with extra steps.

So if you're hearing this, it was me who said yes. Thanks for listening, and if you try one thing after this, pick a task you keep putting off, give an agent the tools to actually do it, and decide in advance where your yes lives.

Fredrik



This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com

Podden och tillhörande omslagsbild på den här sidan tillhör Fredrik Ahlgren. Innehållet i podden är skapat av Fredrik Ahlgren och inte av, eller tillsammans med, Poddtoppen.