Daily Agentic Field Watch - 2026-08-11 21:00 UTC

• MAI-Code-1.1-Flash is now rolling out in GitHub Copilot. Microsoft says the model adds native vision support, improves coding quality/tool use, cuts token usage by 25%, runs 25% faster in Copilot, and is priced at one quarter of MAI-Code-1-Flash; GitHub says it is available across Copilot CLI, cloud agent, app, GitHub, VS Code, Visual Studio, mobile, JetBrains, Eclipse, and Xcode. Sources: Microsoft AI announcement, GitHub Changelog

• Copilot’s measurement stack got a real enterprise bump: the usage metrics API now breaks out agent app activity by individual agent, the impact dashboard adds ROI math (cost/dev/month, % payroll/month, pull requests/month), and the usage report now exposes per-model token breakdowns so admins can trace spend more precisely. Sources: usage metrics API, ROI dashboard, per-model token breakdown

• The August 3 Copilot weekly bundle is a good ergonomics update for agentic work: the app now shows model/cache/credit metadata per request, CLI adds a sessions sidebar plus /worktree and /rewind, and VS Code gets element-level browser feedback, /btw side chats, multilingual dictation, and richer Markdown diff review. Source: GitHub Copilot weekly releases - August 3

• The Copilot SDK and Agent Framework are tightening the runtime contract. The SDK now adds history.clearContext, Tool.isTerminal, and managed permission injection at session start, while the Agent Framework post shows MAF can use the Copilot SDK backend and power multi-agent orchestration like Squad. Sources: Copilot SDK release, Agent Framework: building agent teams with Copilot CLI and Squad

• Semantic Kernel 1.79.0 continues the transition toward Microsoft Agent Framework: the .NET release includes an AF 1.0-compatible migration/samples update, Python docs add a successor callout, and the release also hardens OpenAPI/MCP behavior. Source: Semantic Kernel releases

• Microsoft Research’s latest production-scale paper on Copilot is worth reading: it samples June 2026 traces from 3.2M users, 13M sessions, 761M LLM calls, and 95T tokens, and frames coding agents as a tool-execution workload rather than just chat. Source: Agentic Coding in the Wild

• OpenAI’s new field report on scientific computing argues agents are shifting researchers from implementation to verification/orchestration, based on eight agent-assisted projects in genomics and related scientific software. Source: Scientific computing in the age of agentic AI