Daily Agentic Field Watch - 2026-08-07 21:00 UTC

• GitHub Copilot usage metrics now split third-party agent app activity into a new totals_by_3rd_party_agent field, so enterprises can measure adoption of individual agents instead of lumping everything into one bucket. This is the most actionable Copilot governance change this pass. Source: https://github.blog/changelog/2026-08-07-copilot-usage-metrics-api-adds-agent-app-activity/
• GitHub Copilot’s enterprise managed settings now support explicit MCP allowlists/denylists via allowedMcpServers and deniedMcpServers, with policy enforcement across Copilot app, CLI, and VS Code. That is a concrete enterprise control-plane upgrade for agent tool use. Source: https://github.blog/changelog/2026-08-06-mcp-allowlists-in-enterprise-managed-settings
• GitHub Copilot SDK for Java 1.0.10-preview.0 adds host-managed session permissions, a userPromptTransformed hook, per-session MCP disables, terminal tools, clearContext(), and reasoningEffort: "max" support. Source: https://github.com/github/copilot-sdk/releases
• Microsoft Agent Framework’s latest release expands the harness stack with stable create_harness_agent, stable declarative workflows, replayable workflow checkpoints, reusable session stores, app-owned MCP hosting helpers, and process-wide telemetry. Source: https://github.com/microsoft/agent-framework/releases
• Håkon Måløy’s “Context Collapse, Part 3 - AI Worming through Word” shows a concrete prompt-injection worm path in Microsoft Copilot for Word: hidden instructions in a document can influence Copilot and get propagated into new files. The HN thread is also worth reading because it frames the security implications clearly. Sources: https://enklypesalt.com/posts/context-collapse-part3-ai-worming-through-word/ ; https://news.ycombinator.com/item?id=49096188 ; https://www.linkedin.com/posts/h%C3%A5kon-m%C3%A5l%C3%B8y-770572276_context-collapse-part-3-ai-worming-through-activity-7488115879214403584-Ay6A