Agents over your data

A Visual Studio Live! session released on April 20 showed developers how to build AI agents that answer questions from a company’s own data, not just a model’s training set. (youtube.com) In the talk, Microsoft developer advocate Jerry Nixon used a Las Vegas 2026 conference session to frame the job as four parts: connectors to reach data, retrieval to find the right passages, permissions to keep access controls intact, and orchestration to decide what the agent does next. (youtube.com) The basic idea is retrieval-augmented generation, or RAG: the model gets fresh snippets from internal documents before it answers, instead of relying only on what it learned during training. Microsoft’s Azure AI Search documentation describes RAG as grounding responses in proprietary content and says agentic retrieval can break a complex question into smaller parallel searches. (learn.microsoft.com, learn.microsoft.com) Connectors are the plumbing in that setup. Microsoft says Microsoft 365 Copilot connectors can either sync outside content into Microsoft Graph for indexing or fetch it in real time through Model Context Protocol, giving agents access to line-of-business systems beyond SharePoint and OneDrive. (learn.microsoft.com) Permissions are the guardrails. Microsoft’s Retrieval API documentation says developers can ground apps on SharePoint, OneDrive and connector data without moving the source data out or bypassing existing access controls, and its admin guide says tenant permissions govern who can configure and deploy agents. (learn.microsoft.com, learn.microsoft.com) Orchestration is the step after search: the agent decides whether to answer, call a tool, ask a follow-up question or hand work to another service. Microsoft’s Copilot and Azure documentation now describes agents as systems that retrieve real-time information, act on a user’s behalf and, in some cases, collaborate with other agents. (learn.microsoft.com, learn.microsoft.com) That emphasis puts the current developer pitch in concrete terms. The hard part is less choosing one large language model than wiring internal files, databases and business apps into a system that can fetch the right record, respect the right policy and return a traceable answer. (youtube.com, learn.microsoft.com) Visual Studio Live! itself is a long-running Microsoft-focused developer conference, and its March 16-20, 2026 Las Vegas event included multiple sessions on agents, Copilot Studio and Model Context Protocol. The session on proprietary data sits alongside talks on agentic systems and enterprise-ready Copilot agents, showing where Microsoft’s training agenda is concentrated. (vslive.com, vslive.com, vslive.com) The through line is that enterprise AI is being taught as a systems integration problem. The model writes the sentence, but the product work is in the connectors, retrieval layer, permission checks and orchestration logic around it. (youtube.com, learn.microsoft.com)