I got stuck on this one for a bit but I learned a TON from it. I decided to deploy two workflows based on this project. One was that weekly newsletter I created for #AIS Day 1, so that I get a daily brief on the most recent & relevant AI news every morning. And the second one is a similar lead scraper to what Nate ran in this example, except I'm targeting a different audience with select criteria. This was a lot of fun to work on, and I also created a skill out of it so that it's easier to build out next time 🔥

The substrate is shifting fast. A few moves I'm watching this week, and where my own build sits relative to them. Qdrant just shipped 1.18 — TurboQuant, per-collection metrics, query audit logs, request tracing IDs. This release is shaped for buyers, not just engineers. The kind of observability that moves a procurement team from "we love this" to "compliance signed off." Andre and team are quietly building the most enterprise-ready vector substrate on the market. NVIDIA's agentic stack keeps deepening — NeMo Guardrails, NIM microservices, the AI Blueprints work. Different lane than mine but worth respect; they're solving the GPU-native end of the same problem. Google's pushing hard on Agent2Agent and AgentSpace. The interop angle matters — if agents from different vendors can negotiate with each other, the whole industry levels up. Anthropic just rolled out clearer line-of-sight into parallel agent execution — what's running, what's waiting, what drifted. Operators have been asking for this for two years. They're shipping the inside view, not just the outside performance numbers. Where my own build sits in this picture: I'm building three faces of one engine. A consumer-facing operator layer that I run my own life on. An education-delivery layer shaped for trades, manufacturing, and veterans crossing into technical work. A security layer I won't go deep on, except to say it's the gap most builders are sleepwalking past. Underneath: a multi-agent runtime built on local-first principles. The cloud is a router for specific calls — never the brain. Retrieval is grounded in my own corpus, not the open internet. Governance lives in structure, not in more capability. What I've learned this year that wasn't obvious starting out: The moat isn't the model. It's the data you keep, the warmth you design in, and how the system behaves when the user is tired. The big labs are building the substrate. The integrators are building the value. Both are necessary; neither is sufficient.