AI can generate more code. That is not the bottleneck. The real work is removing integration, ownership, approval, and governance bottlenecks with a domain-driven MCP platform, APIM, and MLflow from mlflow.org.
The useful version of agent memory is not a bigger bucket. It is a layered system: compact truth, raw notes, topic summaries, and a local index that works even when embeddings do not.
Daniel Kokotajlo's AI forecasts look extreme until you separate the mechanism from the calendar. The direction is plausible. The 2027-style timeline is the part I struggle to buy.
The final leg of our 23 days in the States with Janice, Nelly, and Jamie: New York City on foot, big buildings doing free entertainment, and two adults trying to keep the moving parts attached.
The middle leg of our 23 days in the States with Janice, Nelly, and Jamie: the Grand Canyon being unfairly huge, Lake Tahoe changing the pace, San Francisco acting as a doorway, and Hawaii cashing the whole thing in.
The first leg of our 23 days in the States with Janice, Nelly, and Jamie: escaping LAX, letting Palm Springs repair us, and taking Las Vegas exactly as seriously as it deserves.
Model Context Protocol can make agents useful inside large companies, but only if it is wrapped in identity, policy, audit, ownership, and a serious gateway architecture.
A useful personal agent needs more than prompts. It needs a workbench, QA, memory discipline, durable task running, evaluation, and docs that stay honest in CI.
Agents are the easy story. Changing the operating model is the hard one.
First impressions from DTW Ignite 2026 in Copenhagen: agentic AI is all over the agenda, but telecom still has the awkward problem of making the economics work.