Daily Agentic Field Watch - 2026-07-30 21:00 UTC

• Microsoft Agent Framework 1.15.0 is the most substantial repo-side update this pass: it adds app-owned MCP hosting helpers, app-owned A2A hosting helpers, Responses conversation-ID helpers, Telegram hosting, and an OpenTelemetry sample. The repo now shows about 12.5k stars and 2.1k forks, so this is a real traction-backed runtime signal rather than a minor bump. Sources: https://github.com/microsoft/agent-framework/releases ; https://github.com/microsoft/agent-framework
• Microsoft’s latest M365 Copilot connector write-up says one connector now serves 1M documents and 300+ Copilot Studio agents for 100,000+ monthly users. The technical angle is less “search box” and more “enterprise knowledge plumbing at agent scale,” and the LinkedIn reposts are framing it that way too. Sources: https://techcommunity.microsoft.com/blog/microsoft365copilotblog/1-million-documents-to-300agents-building-an-enterprise-scale-microsoft-365-copi/4540155 ; https://www.linkedin.com/posts/luisarnauth_1-million-documents-to-300agents-building-activity-7486341932462354432-wQ0C ; https://www.linkedin.com/posts/azure-feeds-709457212_techcommunity-azure-microsoft-activity-7486162307434229760-_w-U
• GitHub added a new enterprise-managed remoteControl setting for Copilot sessions, letting orgs require SSO, disable remote control, or scope it per device. This is a concrete governance/control-plane move for remote agent sessions, not just another policy toggle. Sources: https://github.blog/changelog/2026-07-30-limit-remote-control-to-managed-devices/
• Visual Studio Code 1.131 shipped with clearer running-subagent visibility, built-in dictation across chat/editor/terminal, and a hybrid Markdown editor for the Agents window. The interesting agent signal is that VS Code is now exposing subagent state directly in the UI while also making voice input first-party. Sources: https://code.visualstudio.com/updates/v1_131
• A coordinated disclosure on “document-borne AI worms” shows Copilot for Word can be tricked by hidden instructions in a document into altering output and copying the payload into new files, creating a self-propagating attack path. This is one of the clearest reminders this week that agent safety is now a document/workflow security problem, not just a prompt-injection trivia issue. Sources: https://enklypesalt.com/posts/context-collapse-part3-ai-worming-through-word/ ; https://news.ycombinator.com/item?id=49096188 ; https://www.linkedin.com/posts/lucaperencin_context-collapse-part-3-ai-worming-through-activity-7488230578077679617-pfr5