GitHub pushed a stronger Copilot observability story today: agent session streaming is now in public preview across cloud agents, CLI, VS Code, Visual Studio, and partner IDEs, and enterprise admins can stream prompts, responses, and tool calls to a SIEM or Microsoft Purview. That makes Copilot sessions auditable instead of opaque. Sources: Copilot agent session streaming is now in public preview
GitHub also tightened Copilot reporting: the usage metrics API now captures CLI suggested lines of code, de-duplicates newer CLI edit counts, surfaces IDE/plugin data for more server-side-visible users, and fixes some AI credit attribution gaps. This is the kind of boring telemetry fix enterprises actually feel. Sources: Improved accuracy and coverage in Copilot usage metrics reports
Copilot CLI can now run in GitHub Actions with the built-in GITHUB_TOKEN, so teams do not need to mint and store long-lived PATs just to automate agentic workflows. Billing can land on the organization directly when the workflow runs in an org-owned repo. Sources: Copilot CLI no longer needs a personal access token in GitHub Actions
Microsoft Foundry posted a practical multi-agent interoperability write-up: Foundry agents can talk to LangGraph agents via A2A, which is a concrete bridge between Microsoft’s enterprise agent stack and a popular open-source orchestration framework. That is more useful than another vague “multi-agent” announcement. Sources: Agent-to-Agent Communication: Connecting Foundry and LangGraph Agents via A2A
Claude Sonnet 5 is now generally available in Microsoft Foundry, and Microsoft/LinkedIn context frames it explicitly around coding agents, research assistants, and business process agents under Azure-native governance. This is a meaningful model availability update for enterprise agent builders, not just a model catalog refresh. Sources: Microsoft Foundry blog ; Yina Arenas on LinkedIn
Hermes Agent shipped v0.18.0, and the release notes are unusually concrete for a fast-moving agent runtime: the repo says P0/P1 backlog is now at zero, Mixture-of-Agents became a first-class selectable model, self-verification landed, and /goal, /learn, and /journey sharpen the long-horizon loop. The project also has real traction, with roughly 208k stars and 37.9k forks, and the Reddit release thread is already active. Sources: Hermes Agent releases ; Hermes Agent repo ; Reddit release thread