Visual Studio brings organization-wide Copilot agents into each eligible repository
Visual Studio now exposes organization-shared Copilot agents inside eligible repositories and lets teams choose reasoning effort by task. Amazon OpenSearch MCP Apps returns interactive charts beside an assistant's text. SMITH trains one policy to create tool schemas and use the resulting tools together. The three systems expand where agent tools are built and used.
Artificial Intelligence··Midday
Shared agents arrive in the repository picker
Visual Studio's August update lets GitHub organizations publish custom Copilot agents across multiple repositories. When an eligible repository opens, the organization's agents appear automatically in the picker, so each repository does not have to carry the same definition separately. Users can also choose low, medium or high reasoning effort by task. The release brings Copilot usage details, Git worktrees and submodule management into Visual Studio. The announcement does not specify which organization plans can use the shared agents.[1]
OpenSearch places an inspectable view beside the answer
Amazon OpenSearch Service now uses MCP Apps to return an interactive view in the chat window beside an agent's text response. Trace waterfalls, service-topology maps and log-pattern displays are generated deterministically from the customer's own OpenSearch data. AWS lists Claude Desktop, VS Code with GitHub Copilot, Cursor, Goose and ChatGPT among supported clients. Setup requires Node.js 22 or newer, credentials with specified read and write permissions, and an OpenSearch UI application connected to the data sources; the announcement gives no regional availability or pricing.[2]
SMITH trains tool writing and invocation together
SMITH trains tool-schema creation and tool use inside one reinforcement-learning policy. After training on 13 procedural-reasoning tasks, a Qwen3 model reached 79.8 per cent macro-average accuracy on held-out tasks. Schema, code and outcome failures received separate rewards, and the researchers report that the generated tools also helped models of other sizes. The study does not test Visual Studio's distribution picker or OpenSearch's chat view. Different explanations and product constraints may govern scientific performance in tool creation, distribution across an organization and a person's inspection of the output.[3], [1], [2]
