In this video I covered three stories that show where AI is heading: more agentic workflows, messier model attribution, and increasingly weird product pressure. First up, Anthropic shipped two big Claude Code upgrades. “Channels” lets me control a live Claude Code session from Telegram or Discord—basically my phone becomes a remote for my coding agent. Under the hood it’s MCP (Model Context Protocol): the channel runs as an MCP server on the same machine as Claude Code, pushes messages into the active session, and my code stays local while the instructions come through a different pipe. They also dropped scheduled recurring cloud tasks. I can hand Claude Code a repo, a schedule, and a prompt, and it runs on Anthropic’s infrastructure—no local terminal babysitting. The examples are exactly what teams actually need: sweeping PRs, analyzing CI failures overnight, syncing docs after merges. Together, channels + scheduled tasks are clearly Anthropic moving toward Claude Code that works in the background and surfaces results whenever I want. Then I broke down the Cursor “Composer 2” controversy. Cursor positioned it like an in-house model, but a developer found the API model ID exposing Kimi K2.5 (Moonshot AI). Access was licensed via Fireworks, but the real issue was disclosure and attribution—especially because Kimi’s license requires UI attribution above a revenue threshold Cursor likely exceeds. Finally, I covered the Wall Street Journal report that OpenAI’s well-being council unanimously voted against an explicit “adult mode,” citing overdependence risks and a ~12% minor/adult misclassification rate—yet OpenAI still appears intent on shipping eventually.