I’ve been playing with AI agent harnesses lately. Basically, what happens between the agent saying “I got this” and me believing it? 😂 🔧 Did the tool finish, or did we just tell it to start? 🔁 If we retry, are we about to send the same email twice? 👻 Can yesterday’s worker wake up and overwrite today’s result? My favorite check is dead simple. Give the agent a file to edit, then have a plain script reopen it and look for the exact change. If it’s missing, the job stays open. Even if the agent wrote a beautiful little victory speech. 🧾 It reminds me of commissioning. “Command sent” and “valve open” get separate lights for a reason. ✅ What’s the dumbest little check that’s saved you from an agent’s victory lap? Bonus points if it was one line of code. 😂

For anyone who has Claude read a log at the start of a session and update it at the end, or is thinking about it: it's a good idea. I use Claude Code, and here's what I added after it bit me. Save as you go, not just at the end. Sessions end early. The moment something is decided or corrected, it goes in the file. Keep "where things stand" separate from "what happened." One short, current file Claude reads every time. One dated history it opens only when needed. One of my project files hit almost 1,500 lines before I split it. Start fresh more often than feels natural. A long chat rereads everything with every message, so usage keeps growing. I once hit 76% of my weekly limit by Wednesday, and my worst chat had accidentally been open for over 37 hours. With a good log, a fresh start costs almost nothing. To set this up, open Claude Code in your project folder and paste this: Set up session logging in this folder. Do all of it yourself, and don't ask me to create anything. 1. Look at what's here first. Never overwrite or delete an existing file. If CLAUDE.md, CONTEXT.md, or LOG.md already exists, add to it. 2. If CONTEXT.md is missing, create it. Keep it under 30 lines, with these sections: Goal, Where things stand, Decisions and corrections, Next step. Fill in what you can tell from this folder and mark anything uncertain "(guess)". 3. If LOG.md is missing, create it for dated history, newest first. If an existing file has old dated entries that belong in LOG.md, show me what you'd move and wait for my yes before moving anything. 4. Add this section to CLAUDE.md, creating the file if needed: ## Session rules - At the start of each session, read CONTEXT.md and tell me in two lines where we left off and what's next. - Save as you go. The moment we decide something, or I correct you, update CONTEXT.md. Don't wait for the end of the session.

A private conversation today got me thinking, so I'm asking the room. When is writing authentic? When is AI-assisted writing authentic? And who sets the standard? What we learned in school, the latest teaching guidelines, Schulz von Thun, Waldorf? If I write in broken English, is that authentic? If I only talk about myself and never ask questions? Or if I do ask, because I actually want to know what you think? I'm autistic. Communicating with people is hard for me to begin with. I use structure to bring order into my thoughts, and I keep it consistent on purpose. Here's how I actually work. I work on a pig farm. Sometimes I'm delivering piglets with my arm inside a sow while I answer your comments. Sometimes I'm on the tractor for hours with nothing else to do. So I dictate everything by voice. I see who commented what, I answer in German, and my AI assistant, which I built for exactly this, translates and structures it so everyone can follow. For me, 2026 is great, because I finally have the freedom to get thirty, forty years of thoughts out of my head. I built a voice box that talks like KITT, better than the original. I built my own OS, the one my grandpa and I dreamed up when I got my first computer. I'm building my own community as a knowledge archive, and something I call Repothek for this community. All at the same time, because my brain is fast. That's what AI is for, to me: accessibility. Freeing up our thoughts. Should I type every reply by hand instead, one by one, with my arm in a sow? How I talk is what we learn in school here in Germany. You ask, you respond constructively, and you ask another question if you want the conversation to go on. Most people say "good morning, how are you?" and then nothing follows, because they don't care. I do care. I remember what you write, and your answers help me revise stuck ideas from my own bubble. So now I'm told I'm not authentic, and honestly, it confuses me. My take: there is no shared language for authenticity. Everyone decides for themselves how they are. Others can accept it or not, and nobody has to join a conversation with me. I don't tell you your face is ugly and ask for a different photo either.

I have been building an operating system that runs our business. I work from a single approval queue with approve, edit, or ignore buttons. I have been learning Jake's ICM structure and had Claude Code analyze and compare my Operating System to Jake's ICM Architecture github skill. In summary, this is what it said about the similarities and differences: "The two share the AI half (instructions in files, a declared list of what each step reads, a check after every step). They differ on what drives the work: in ICM, folders and an agent session do the orchestration, while your OS uses code and a database." I am leaning towards continuing to build out my Operating System as is, because my team still uses the connectors (like Go High Level for leads and Clickup for project management) and Jake's ICM would have us using folders for each client. That would be pretty disruptive to our ops. Is this a problem? I am not a developer, just an entrepreneur who is learning how to harness AI, so a lot of this can go over my head. If you want see the full comparison that Claude gave me, I am attaching a photo below.

Hi all 👋 Thank you for all the great content. I’ve been lurking around for a few weeks and it’s been a lot of fun to try new things. Unfortunately I can’t share any of it but I’m hoping to try one of the competitions soon. Most of my ICMs now have evolved to have roughly 3 rooms: 1. Data extraction and pre-processing 2. Data analysis and feature extraction 3. Result presentation and discussion The projects are large enough that it is starting to feel like each room needs its own coding subagent because re-usable skills don’t seem to cut it anymore: 1. Engineers : each data source gets its own CLI, each analysis implements a different research angle, each user interfaces have different taste requirements… 2. Reviewers : Looks for red flags not just in the code but also in the data and the rendering of the presentation… 3. Executors : Runs my daily pipeline fast and without polluting the context window. Most of the generic skills I find online are for software engineering; not data science. I used to think specs folders and test folders should sit at the root of my repo but now I think they should be buried deep into each room? Anyone else trying to use ICM for data science? Have you used ICM to migrate away from Jupyter notebooks or do you combine Jupyter with ICM? I’m also curious if you have a particular way of forcing your ICM to invoke sub-agents and restrict their tools and/or folder access. I’m not doing any of that right now… Cheers! E