Microsoft Research launched Orchard, an open-source agentic modeling framework centered on Orchard Env, a reusable Kubernetes-native environment service for training and evaluating agents across SWE, browser, computer-use, and personal-assistant tasks. The repo is already visible but still small-scale traction-wise at 114 stars and 6 forks, and LinkedIn commentary correctly latched onto the environment layer as the bottleneck. Sources: Microsoft Research blog ; GitHub repo ; LinkedIn discussion
Microsoft released Resource2Skill, which turns tutorials, docs, repos, and other human-created resources into executable skills and ships both a runtime and a Skill Wiki. The GitHub repo shows 351 stars and 33 forks, the Hugging Face dataset release adds the released skill libraries, and LinkedIn notes the paper’s claim that the same agent beat Claude Code/Codex on 26 of 28 tests. Sources: GitHub repo ; Hugging Face dataset ; LinkedIn breakdown
Microsoft Copilot Studio rolled out the new GitHub Copilot harness for long-running business processes. Microsoft says the new harness is aimed at more complex, multi-step work, the LinkedIn discussion frames it as giving Copilot Studio agents the coding/reasoning power behind Copilot Cowork and the GitHub Copilot coding agent, and the Reddit thread is already asking whether evals and the test pane now consume credits. Sources: Microsoft Community Hub ; LinkedIn discussion ; Reddit thread
Microsoft’s FY26 Q4 earnings call is the clearest Satya-linked strategy signal this pass: the transcript says Copilot is moving from chat to Cowork to Autopilots, that Autopilots are autonomous long-running agents with enterprise compliance, and that Microsoft will bring Code together with the rest in one super app across consumer and commercial experiences. It also says the company is separating the harness, context, memory, and action space from any one model family. Source: Microsoft Investor Relations transcript