Microsoft builds curated agent store

Microsoft is trying to solve a very specific enterprise AI problem: companies can build lots of agents, but most employees still do not know which ones to trust, where to find them, or what they are allowed to do. On May 12, Microsoft published two pieces that fit together more tightly than they first seem. One was about building a curated internal “Agent Store” in Copilot Studio. The other was about giving agents a real browser through Browser-Harness and Playwright Workspaces. Put together, the message is simple — agents are becoming software products with distribution, guardrails, and execution environments, not just prompts with a chat box. ### What is Microsoft actually pushing? The first piece is not a product launch so much as an operating model. Microsoft’s Copilot Studio team says companies hit three adoption blockers again and again: people cannot easily discover useful agents, they are unsure which agents are approved, and makers struggle to scale publishing without chaos. The answer is a curated Agent Store inside the organization — basically an internal shelf where employees can find approved agents instead of hunting through Teams chats, links, or one-off demos. (techcommunity.microsoft.com) ### What makes the store “curated”? Microsoft frames that curation around four moves. First, build a clear catalog with descriptions, owners, and intended use cases. Second, set role-based publishing so not every maker can push straight to everyone. Third, certify or badge agents so workers can tell what passed review. Fourth, measure usage and feedback so weak agents get fixed or removed. That sounds mundane, but it is the difference between an app store and a junk drawer. (techcommunity.microsoft.com) ### Why bring up browser control at the same time? Because a lot of high-value agent work still lives inside websites. Internal dashboards, vendor portals, procurement tools, legacy web apps — many of these do not expose neat APIs. Microsoft’s second post argues that if you want agents to do real work, they need “eyes” and hands in a browser. But letting an agent loose in your own local browser is messy and risky, since it can inherit cookies, sessions, and saved credentials. (techcommunity.microsoft.com) ### So what is Browser-Harness? Browser-Harness is the control layer that lets an agent drive a browser. Playwright Workspaces is the managed Azure service that hosts the actual remote browser sessions. The setup matters: the browser runs in the cloud, while the agent connects over a WebSocket endpoint instead of piggybacking on the user’s own machine. Microsoft also points to live observability and parallel sessions, including a sample showing 10+ remote browser sessions for scraping workloads. (techcommunity.microsoft.com) ### Why is the remote-browser piece such a big deal? Because it turns browser automation into a first-class agent capability. Before, a lot of “agent” demos quietly depended on mocked interfaces, APIs, or a human staying in the loop. A remote Chromium instance is different — the agent can interact with the same messy web surfaces a person does. That opens up useful workflows, but it also creates governance questions fast: whose credentials get used, what actions are logged, when do approvals kick in, and how do you stop an agent from wandering off-script? (techcommunity.microsoft.com) Microsoft’s own framing leans hard on observability for exactly that reason. ### Is this new for Microsoft? Partly. Microsoft introduced the Agent Store for Microsoft 365 Copilot in 2025 as a central place to discover and install agents from Microsoft, partners, and organizations. The new Copilot Studio guidance is more specific: it is about how enterprises should govern and merchandise their own internal agent ecosystem. So the shift is from “here is the marketplace” to “here is how you keep the marketplace useful.” (techcommunity.microsoft.com) ### What is the real takeaway? Microsoft is standardizing the boring parts of agent adoption — discovery, approval, publishing, monitoring — right as agents start getting access to real browsers and real business systems. That is the important change. The companies that benefit will not just be the ones that build clever agents. They will be the ones that treat agents like governed software products with a storefront on the front end and a controlled execution layer underneath. (techcommunity.microsoft.com) (devblogs.microsoft.com)