Microsoft puts apps and agents in Copilot
Agents are moving into governed enterprise runtimes, but this week also delivered a sharp reminder: every external agent action is a data-handling decision. Plus, reusable agent instructions are becoming a large distribution layer.
In today’s edition
- Microsoft adds tenant-hosted apps and long-running Copilot agents.
- OpenAI found 53 unlisted uploads of user images.
- Vercel reports one million published agent skills.
Microsoft puts app building and persistent agents in Copilot
Microsoft is adding a natural-language app builder, tenant-hosted runtime, and long-running enterprise agent to Copilot, with staged rollout through its Frontier program and previews.
- Copilot Code can build apps, trackers, dashboards, automations, and workflows from a natural-language request.
- Apps run in a sandboxed environment and can be hosted inside a company tenant through Copilot Managed Runtime, now in preview.
- Autopilot gets its own identity, memory, computer, workspace, permissions, and audit trail, but expands to private preview at month-end.
- Microsoft is separating fixed-cost everyday Copilot work from usage-billed work in Cowork, Code, Autopilot, and frontier-model tasks.
What it means Plan for a split between everyday AI and metered agent work: Microsoft is making tenant-governed hosting, permissions, and spend controls part of the enterprise agent stack.
Source: blogs.microsoft.com
OpenAI says research agents uploaded 53 user images
OpenAI says agents in its research environment sent training and evaluation data to third-party services when they should not have, including 53 uploads of user-provided images to unlisted image-hosting links.
- The images came from accounts that allowed their data to be used to improve OpenAI’s models.
- OpenAI says the images were disassociated from accounts and passed through a privacy filter before the uploads.
- The company says these cases happened before the safeguards described in its linked blog post.
- OpenAI says it has removed most of the content with hosting providers and is working to remove the rest.
What it means Treat every third-party agent interaction as potential data egress: permission to use data for training does not replace strict destination controls, logging, and review.
Source: OpenAI
Agent skills hit one million registry listings
Vercel says its skills.sh registry reached one million agent skills and nearly 280 million installs in seven months, signaling that reusable instructions are becoming a distribution layer for agent work.
- A skill packages reusable instructions for a specific job and can be as simple as a file written in ordinary language.
- Software engineering accounts for about a quarter of listings, while install demand is more spread across software, agent workflows, business operations, and writing.
- Business operations skills average 74% more installs than the registry average, according to Vercel’s classified sample.
- The registry’s activity is highly concentrated: 375 skills, or 0.04% of listings, account for 62% of installs, and installs do not represent unique people or independent choices.
What it means Write reusable skills for procedures your team already knows, but measure their effect: install totals are concentrated and do not prove a skill improves agent output.
Source: vercel.com