Microsoft is pushing a single enterprise agent system story instead of a pile of point products: Agent Factory is now framed as the consumption layer spanning Microsoft 365 Copilot, GitHub Copilot, Fabric, Foundry, and agents built elsewhere, while Agent 365 is the control plane for observe/govern/manage/secure plus spend tracking. The same June posts emphasize model diversity, Microsoft IQ grounding, and continuous improvement loops driven by evals and human feedback. Sources: Achieving success with AI ; AI alone won’t change your business. The system running it will. ; AI, jobs, and the next generation
Microsoft is formalizing agent skills as a distributable layer: microsoft/skills is now a 2.6k-star repo with 297 forks and 175 skills, while microsoft/skills-for-fabric is a new Fabric-specific bundle with 610 stars and 160 forks. The June Power BI update says report-authoring agent skills can plan, design, author, validate, and publish reports, with the Desktop Bridge enabling a continuous edit-verify loop. Sources: microsoft/skills ; microsoft/skills-for-fabric ; Power BI June 2026 Feature Summary
Copilot Notebooks got a more visual workflow: mind maps are GA worldwide, turning notebook content into explorable branches instead of a flat Q&A flow. The June notebooks update and the PM’s Reddit post both note that the maps are private to the user and retained for 30 days, and that the feature is now reaching Copilot Chat users too. Sources: What’s New in Notebooks | June 2026 ; Reddit discussion
GitHub Copilot app v1.0.3 is a meaningful ops/stability release: it auto-recovers from a critical empty/corrupt database bug on startup, adds multi-hour daily schedules, creates a dedicated Sessions tab, and shows a full context-window breakdown in the session info popover. Scott Hanselman’s LinkedIn notes are useful field feedback on the same release cycle, especially around remote control and scheduled workflows becoming real workflow primitives. Sources: GitHub app v1.0.3 release ; Scott Hanselman on LinkedIn
The reliability conversation is getting more serious on both the architecture and evaluation sides: Martin Fowler’s Building Reliable Agentic AI Systems case study leans on retrieval, a reflection agent, citations, and daily evals, while a new arXiv preprint argues that current coding benchmarks are misaligned with agentic software engineering because they collapse model, harness, and environment into one score. Sources: HN discussion ; arXiv preprint