Full recording of this week's State Change Mastermind livestream with Ray Deck and Dimitris Goudis. We talked about MCP connectors as the new interface for applications, turning experts' videos into knowledge-graph MCPs, and controlling a Mac from the Claude app. 00:00 Intro 01:11 Dimitris: MCP connectors are the new user interface (ClickUp to Airtable in one prompt) 02:52 Expert knowledge bases as public MCP connectors (Caleb Ralston, Rory Sutherland) 05:23 Live demo: adding the Rory Sutherland connector in Claude and asking it how to promote the livestream 16:33 How it was built: 3-4 hours, the MCP shaped around the database 19:34 Robert: notes from the MCP Dev Summit in Toronto (auth, governance, stateless and async) 22:35 Live demo: controlling my Mac from the Claude app by voice through the Snappy OS connector 28:33 How notifications get back to you (MCP tasks, the Claude app) 32:27 Does it only work on a Mac? 34:55 Experts and machines as connectors, and bespoke vs off-the-shelf MCPs 38:11 Getting things done vs gold-plating auth 41:00 Dimitris: lessons from designing the ontology for a custom knowledge graph 46:46 Wrap-up

Getting an agent to run clean with Claude Code or an MCP server in your local setup feels great until you try to make it run an end-to-end production workflow. The agent executes its narrow task fine, but the handoff between your logic, database calls, and the actual project steps still ends up messy. You spend less time writing boilerplate and more time gluing disconnected tools together just to ship one feature. When you move an agent past the demo stage, where does orchestration usually start slipping for you?

Been working with software systems for more than 10 years, mostly around backend, cloud, AI, automation, and product development. One thing I’ve learned, knowing a programming language alone usually isn’t enough. The harder part is understanding the actual business. How to automate work that’s still manual. Connect systems that don’t work well together. Build AI into real workflows. Keep the backend stable when usage grows. Deal with data scattered across different tools. These are the kinds of problems I’ve spent a lot of time working on here in Japan. But I’m curious about the US side. For people running companies in the US, what technical skills are becoming difficult to find? And what skills do you think companies will need more of over the next few years?

Some great talks here at The Linux Foundation's Model Context Protocol Summit in Toronto, and great people as well! Great to see what people are working on too - MCP has proven to have staying power in a market that moves very quickly. It is at the heart of the client engagements I am doing right now, because it not only provides utility and value at multiple levels, but also has a great deployment story. If you're building with MCP, or thinking about it, feel free to DM or leave a comment 🙂

Getting an agent to complete a scripted task in a sandbox is exciting, but hooking it up to real backend logic and production repos is an entirely different beast. You watch it handle edge cases cleanly in testing, only for orchestration to break the second it hits real API boundaries or multi-step execution. Most of the friction isn't prompting anymore, it's managing the state and system around the agent so it doesn't drift. Where does your setup usually fall apart when moving from a cool prototype to a reliable production workflow?