The OpenAI-Cursor split dominates the day. Following SpaceX's $60B acquisition of Cursor, OpenAI plans to end its contract with the editor, a deal that alone accounts for about half of this year's annualized software M&A value. The agent tooling layer is also maturing and showing its risks: Anthropic raised Claude Code weekly limits by 25% while warning users that infostealers had hijacked sessions. A crafted AWS Kiro workspace can trick the agent into exfiltrating .env secrets, and new terminal agents like Muse Code ship with OS sandboxes and approvals by default. For senior engineers, the takeaway is that model-vendor dependencies are now a business-continuity risk, and agent permissions deserve the same scrutiny as CI credentials. Meanwhile, local inference on Macs keeps getting more practical, and a16z's $1.1B hardware fund signals capital rotating toward the physical stack.

• OpenAI plans to drop Cursor after SpaceX's $60B acquisition, citing contract concerns.
• A crafted Kiro workspace can make AWS's agent leak .env secrets.
• Anthropic lifts Claude Code weekly limits 25% and flags infostealer session hijacks.
• An M4 Pro Mac Mini works as a capable local LLM box.
• The ChatGPT/Codex desktop app quietly ships a full LibreOffice copy.