Satya Nadella’s June 2 LinkedIn post, “How do we build a frontier intelligence ecosystem?”, is the clearest strategy signal this pass: Microsoft is framing the stack as Windows + IQ/Web IQ + autopilots + 7 new MAI models, with autonomous long-running agents positioned as enterprise-grade “claws.” Source: LinkedIn.
Microsoft Foundry’s Build 2026 agent update is the biggest platform move here: the company is now explicitly describing a build-in-GitHub, deploy-in-Foundry, operate-with-tracing/evals loop. Hosted agents run in per-session sandboxes, support long-running routines, and are framework-agnostic; GA is said to be within the next 30 days. Sources: Build 2026 post and hosted agents / enterprise distribution.
Foundry’s enterprise distribution story got sharper with autopilot agents: they get their own identity plus email, calendar, OneDrive, Teams access, and org-chart presence, and the workflow is explicitly “goal -> ongoing execution -> checkpoints -> collaboration” rather than prompt/response. Sources: enterprise distribution post and workstream manager sample.
Microsoft is also tightening the commercial and FinOps story. Judson Althoff’s June 16 blog says Microsoft Agent Factory is the single consumption model across Microsoft 365 Copilot (including Cowork), GitHub Copilot, Fabric, Foundry, and Copilot Studio, while Agent 365 is the control plane for governance and cost management. Source: Achieving success with AI.
Semantic Kernel continues to ship like an active platform, not a museum piece: the repo is at 28.2k stars, 4.7k forks, 272 releases, and python-1.43.1 landed today, June 17, 2026. Sources: repo and releases.
Anthropic’s new Claude Code research says agentic coding is settling into a division of labor: from ~400k sessions, humans make about 70% of planning decisions while Claude makes about 80% of execution decisions, and success tracks task expertise more than job title. HN is already discussing it. Sources: Anthropic research and HN thread.
PwC’s “Is Grep All You Need? How Agent Harnesses Reshape Agentic Search” is the freshest non-vendor research signal worth watching: harness design may matter as much as the retrieval method, and Jerry Liu’s LinkedIn summary notes grep outperformed semantic search in their setup. Caveat: the study is on conversational memory, not code corpora. Sources: arXiv, HN thread, and LinkedIn discussion.