Daily Agentic Field Watch - 2026-07-07 21:00 UTC

• Microsoft Agent Framework moved Agent Skills for .NET out of experimental preview today. The stable API matters because it turns skills into a production-ready packaging/governance layer for reusable domain expertise, and the repo still has real momentum at about 11.9k stars and 2k forks. Sources: https://devblogs.microsoft.com/agent-framework/agent-skills-for-net-is-now-released/ ; https://github.com/microsoft/agent-framework
• Microsoft Research published two July papers that feel especially relevant to agent infra: one argues web agents should use typed actions / Web Verbs instead of click-heavy browsing, and another introduces MOSAIC, a plan-check-act-or-refuse loop for safer multi-step tool use. Together they point to a more auditable, contract-driven agent stack. Sources: https://www.microsoft.com/en-us/research/publication/web-agents-should-use-typed-actions-instead-of-click-based-browsing/ ; https://www.microsoft.com/en-us/research/publication/learning-when-to-act-or-refuse-guarding-agentic-reasoning-models-for-safe-multi-step-tool-use/
• GitHub Copilot CLI pushed a new prerelease today, 1.0.69-3, with a small but telling governance tweak: built-in file edits can bypass the sandbox once you approve it. The repo is also clearly alive at roughly 10.9k stars, 1.7k forks, and 344 releases. Sources: https://github.com/github/copilot-cli/releases ; https://github.com/github/copilot-cli
• OpenAI Agents SDK landed v0.18.0 today. The repo is still one of the largest in the category at about 27.7k stars and 4.3k forks, and the README keeps emphasizing the full harness stack: sandbox agents, handoffs, guardrails, sessions, tracing, and realtime agents. Sources: https://github.com/openai/openai-agents-python
• Anthropic’s new Claude Code usage study is one of the better adoption signals this week: coding-agent activity on GitHub has more than doubled since late 2025, users average about 20 hours per week, and the report says people still own the “what” while the agent increasingly owns the “how.” Source: https://www.anthropic.com/research/claude-code-expertise
• Satya Nadella’s new essay, surfaced by Microsoft Signal, is worth the read for the strategy layer: Microsoft is arguing that firms need to compound human capital and token capital through a durable learning loop, not just swap models. That frames agents as organizational memory and IP infrastructure, not a chat feature. Sources: https://snscratchpad.com/posts/frontier-ecosystem/ ; https://news.microsoft.com/signal/home/