Microsoft Agent Framework dotnet-1.17.0 is the freshest Microsoft-side runtime release: Durable Task and Azure Functions integrations were split out of core, Handoff sample responsiveness was fixed, and declarative workflows now fail fast when an agent returns an error. Source: GitHub release
Microsoft’s Windows on-device agent story got more concrete in the Build-era Windows developer material: Aion 1.0 Plan is described as a 14B reasoning and tool-calling model meant to ship in-box on capable Windows devices, enabling local workflows that can reason over intent, invoke tools, manage files, and orchestrate sub-agents. Sources: Windows developer blog | Build26 news repo
Microsoft Research’s ENCO paper describes a production-scale engineering copilot with a lightweight hierarchical agentic planner and NL2SearchQuery retrieval, and says it has been live since Sep 2023 with tens of thousands of interactions and 1,000+ monthly active users. Sources: Microsoft Research | ACM publication
Microsoft Research’s WorkflowPerturb adds a useful calibration benchmark for multi-agent workflow metrics, with 4,973 golden workflows and 44,757 perturbed variants spanning missing steps, compressed steps, and description changes. Sources: Microsoft Research | arXiv
OpenAI’s GPT-5.6 launch is a meaningful tool-use update: the post highlights improved long-horizon knowledge work, stronger browse/tool/computer-use performance, and better handling of messy workflows across productivity apps. Source: OpenAI
OpenAI’s GPT-Red is the week’s biggest agent-security signal: it trains an automated red-teamer through self-play, outperforms humans on a prompt-injection arena, and is used to harden GPT-5.6 against malicious tool-use scenarios. Source: OpenAI
GitHub Copilot is deprecating several models on September 1, 2026, including Gemini 3.1 Pro, Claude Opus 4.5/4.6, Claude Sonnet 4.5/4.6, and Raptor Mini. That matters for agent workflows and model-policy pinning in enterprise setups. Source: GitHub Changelog