Consider this a thank you @Jake Van Clief for the inspiration and giving us the mindset and tools to grow our businesses. ***Disclaimer: this is an advanced setup. DONOT use until you're fully comfortable with Jake's method*** =============================== Many of us eventually hit the limit with basic chat interfaces. The agent forgets context between sessions, data security starts to feel risky, and the setup creates more friction than value. This stack addresses those issues directly It combines Hermes Agent with Cognee as the memory engine, hosted on a simple DigitalOcean Droplet and secured through Cloudflare Tunnel. The entire deployment follows Jake Van Clief’s Interpretable Context Methodology for clean, repeatable orchestration. The result is a private, self-improving AI agent that grows more capable over time while keeping your data and server fully under your control. The components stay minimal and transparent: Hermes Agent as the autonomous gateway, Cognee for structured relational memory, and Cloudflare Tunnel for secure outbound-only access. You deploy once using the ICM workflow, then the system handles the repetitive memory management and self-improvement loops. The agent becomes a genuine thinking partner instead of a one-off responder. That frees up your attention for the judgment and creative work only you can do. If you are running a self-hosted agent setup or exploring similar private stacks, I would like to hear what you are using and what friction you have solved. Drop your thoughts below.

First Monday. First winner. Let's go. @Alexander Paschka topped the 7-day leaderboard with +454 points and earned himself free lifetime VIP access. (Thousands of dollars in value) He was already a paying VIP member, so we upgraded him to Free lifetime access. Free members get upgraded to Premium etc. That's how this works. You show up, you lead, you get taken care of. Now here's the part worth paying attention to. Alexander didn't rack up 454 points by gaming the system. He posted a real Win🏆. One of his clients is buying 10 seats of Claude Enterprise, and Alexander is leading the training. He learned how to do that here. Then he went out and sold it. The community went crazy because that post was useful. People wanted to know how he did it, what he said, how the conversation went. That's engagement you can't fake. You share something real, people respond. And that's the lesson buried in the leaderboard. The thing that got Alexander to #1 is the same thing that works on LinkedIn, on YouTube, on any platform where attention matters. You give people something they can use and they come back for more. The skills that make you valuable in this community make you valuable everywhere else too. I mean its the reason this community exists in the first place. And now you have a place to practice it, if it works here you know it will work every where else. Shoutout to @David Vogel at #2 with +221. David has been showing up in almost every post and every live class adding real, specific, helpful input. Consistent. Week after week. (He earned VIP early on) Keep watching that name. Millenial Cat and Shirsho Guha rounded out the top 4. Respect. The leaderboard resets soon. New week. New race. Whoever is on top next Monday morning gets the same deal. Free lifetime Premium if you're on the free plan. Free lifetime VIP if you're already VIP.

Hola, For the past few days, my timeline on X/twittet has been nothing but people canceling their subscription and reporting that 4.6 has been dumber/nerd with many rolling back to 4.5. or moving to Codex or something else. Thankfully the file system doesn’t care what model you use but wondering if folks here have noticed the same thing (if using Claude) ? I’ll say that my usage in Claude at work has been okay, didn’t put too much though about it since we have unlimited usage at the moment but will be testing this week and seeing if it makes more mistakes.

Is the projects folder on claude similar to creating a folder via the terminal? I know this is elementary but before I knew about Claude code, I created a project to be a "sherpa/ghostwriter" asking me introspective questions, trained on my writing style, writing in general, and we went chapter by chapter speaking out the book. Would it work better in code? I know the folders on the computer allow AI to maintain the memory but is that different in how projects on claude work?

For those of you who've been following me on social media for a while, you've heard me say this over and over: building your own agent frameworks is a waste of time. Stop doing it. Stop paying people to do it. Stop learning how to do it. (Unless you like local projects, that's fun) Anthropic just published this today: https://www.anthropic.com/engineering/managed-agents Here's what it is in plain terms. Managed Agents is a hosted service where Anthropic runs the entire agent infrastructure for you. The loop that calls Claude, the sandbox where code runs, the context management, crash recovery, security, scaling. All of it. You define what the agent should do (system prompt, tools, connections to your systems) and they handle the rest. The key line from their own engineering team: "Harnesses encode assumptions about what Claude can't do on its own. Those assumptions need to be frequently questioned because they go stale as models improve." They gave a specific example. They built a workaround into their harness because Sonnet 4.5 would quit tasks early when it sensed its context limit approaching. When they ran the same harness on Opus 4.5, the problem was gone. The fix became dead weight. One model release made their own engineering work obsolete. Now think about what that means for every startup and every freelancer building custom agent harnesses and selling them to clients. Every assumption they baked into their code is a bet against the next model release. And the frontier labs are shipping new models faster than anyone can maintain a harness. This is the thesis. Your value lives above whatever just got commoditized. The infrastructure layer of agents just got commoditized. The thinking about what to build, why to build it, and how to structure the work around it did not. That's what we teach here. That's why we focus on the 60/30/10 framework, on understanding which layer a problem belongs on, on prompt architecture and workflow design.