Daily Agentic Field Watch - 2026-09-03 09:00 UTC

• GitHub Copilot’s enterprise controls tightened again: admins can now set any default model for new conversations, Copilot app and CLI now honor content exclusion policies, and Copilot code review can approve pull requests in public preview. The PR-approval change already has HN discussion; the model-default and exclusions updates are the sharper governance signals. Sources: enterprise-managed default model | content exclusions GA | code review approvals | HN thread

• GitHub Copilot CLI v1.0.83-2 and v1.0.83-1 shipped useful agent ergonomics: custom agents can try several models in order until one is available, model-policy: required keeps the list pinned, claude-fable-5.1 is supported, Linux sandboxes now restrict egress to the configured proxy, and the prior patch added session sorting, forceLoginOrgs, and moved /mcp config into the plugins dashboard. Source: Copilot CLI releases

• GitHub Copilot SDK v1.0.13-preview.4 added rewind/file-change tracking across SDKs, session-scoped token providers, trusted built-in plugin directories, permission decisionContext, sensitive env-var requests for Node extensions, argsSchema validation, and Java in-process runtime support. The repo itself has real traction, with 10.5k stars and 1.4k forks. Sources: SDK releases | SDK repo

• Microsoft Agent Framework introduced agent and workflow channels in Python for OpenAI Responses, Telegram, A2A, and MCP, with shared hosting/state and protocol-specific adapters so the app owns routing, auth, storage, and deployment while the framework handles conversion and execution. That is a meaningful step toward multi-channel production agents instead of one-off demos. Source: Introducing agent and workflow channels

• Microsoft Copilot Studio now documents the three harnesses clearly: the GitHub Copilot harness for reasoning-heavy, multi-step work, the standard harness for rule-based flows, and the Copilot chat harness for extending Microsoft 365 Copilot Chat. The billing doc also makes the cost model explicit: GitHub Copilot harness usage is credit-based from the moment you start building, including model, tool, and harness consumption. Sources: harness overview | usage-based billing

• Windows 365 for Agents got a dedicated Intune security baseline generally available, plus agent-pool device grouping and Autopilot preparation profile support in preview. That is a concrete step toward treating agent Cloud PCs like governed infrastructure instead of disposable demos. Sources: Windows 365 for Agents what’s new | Intune baseline settings