Microsoft’s September 2026 SharePoint Copilot roundup pushes the product from “ask questions” toward reusable agent tooling: skills now follow users across SharePoint and OneDrive, Copilot can evaluate those skills, there’s guidance for SharePoint-safe HTML generation, image understanding across document libraries, and workflow actions like Teams notifications and approvals. Sources: original announcement ; Julie Seto’s LinkedIn context
OpenAI’s Astra safety pre-read plus safety overview is the biggest agent-safety update this pass. OpenAI says Astra is its first model to hit the Critical cybersecurity threshold, has added trajectory-level monitoring and stronger misalignment controls, and notes that monitorability drops under adversarial conditions. Sources: Path to Astra ; Safety overview ; HN discussion ; HN system-card thread ; Reddit discussion
OpenAI’s “Research acceleration: The view inside OpenAI” is a strong adoption signal for coding agents in real work: inside OpenAI, researchers are now using agents daily, median usage is over $600/day of inference, the 90th percentile is above $7k/day, and the organization is spending 3.1 agent-workdays for every human workday. Sources: OpenAI post ; HN discussion