Daily Agentic Field Watch - 2026-09-23 20:00 UTC

• Microsoft Agent Framework added Foundry hosted-agent isolation controls that separate user-data isolation from session isolation for code/files, supporting direct callers, trusted middle tiers, app-managed sessions, and shared session pools. This is a concrete enterprise boundary model for long-lived agents, not just a hosting toggle. Source: Microsoft Agent Framework

• Microsoft Agent Framework is moving specialist-agent expertise toward distributed skills over MCP: the orchestrator can keep a single model while specialist instructions and tools remain in domain services, reducing duplicated agent deployments and making skills independently maintainable. Source: Microsoft Agent Framework

• GitHub Copilot app local sandboxing entered public preview. Per-project policies can constrain filesystem access, outbound/local-network access, and which GitHub/Git credentials are exposed; if the OS cannot enforce the requested policy, the shell fails closed. It is off by default and separate from cloud/remote sandboxing. Source: GitHub Changelog ; Reddit discussion

• GitHub Copilot app added enterprise-managed OpenTelemetry export for agent sessions, model requests, and tool activity, with prompt/response content excluded by default. This gives operators a standard trace surface for debugging and governance across agent runs. Source: GitHub Changelog

• GitHub documented a production-scale Copilot runtime rewrite from TypeScript/Node to more than 800,000 lines of Rust, delivered incrementally through 128 pull requests with agents writing most of the code. GitHub says the new runtime is lighter, embeddable across six SDK languages, and materially faster; the HN discussion is treating it as a notable example of agents making a previously uneconomic rewrite feasible. Sources: GitHub engineering post ; HN discussion

• Anthropic released Claude Opus 5.5, positioning it as a cheaper, faster Opus-tier model for long-running agents and coding: the official page says it matches Fable 5.1 on most work, costs 40% less than Opus 5, and is priced at $4/M input and $20/M output tokens, with faster mode available. Independent community reports are strongly positive on multi-session coding and token efficiency, but remain anecdotal and early. Sources: Anthropic announcement ; GitHub Copilot availability ; Reddit practitioner reports

• GitHub Copilot expanded OpenAI’s GPT-6 family with GPT-6 Sol, a balanced interactive/agentic-coding model, and GPT-6 Luna, a lightweight low-cost option, both rolling out across Copilot surfaces with usage-based billing. The HN and Reddit response already frames Sol as a workhorse candidate and Luna as a routing/price option; those comparisons are community-reported, not an independent benchmark. Sources: GitHub Changelog ; HN discussion ; Reddit discussion

• xAI’s Grok 4.7 began rolling out in GitHub Copilot for agentic coding and complex multistep workflows. GitHub exposes it across IDEs, CLI, cloud agent, app, and enterprise plans with provider-list usage billing; early Reddit coding reports are mixed, so capability claims should be treated as official positioning plus unverified practitioner signal. Sources: GitHub Changelog ; Reddit discussion

• Microsoft 365’s federated Copilot connectors are now documented around MCP-backed external data access with centralized tenant controls, including an October 20 transition from the deprecated CLI toggle to Allowed agent types. The Work IQ SharePoint MCP connector is in preview across Copilot Studio, Power Apps, Power Automate, and Logic Apps; a practitioner thread reports the practical effect as centrally governed access to services such as GitHub, WordPress, and Salesforce. Sources: Microsoft Learn: federated connector governance ; Work IQ SharePoint MCP ; Reddit discussion