Two official posts land on October 28, 2026, with different clocks and different words for availability. The Microsoft blog is timestamped 16:51:35 UTC and carries the byline of Tina Schuchman and Vladimir Fedorov. The Foundry post is timestamped 16:45:06 UTC. The Microsoft blog says the Microsoft Copilot harness is being made available through Microsoft Foundry on that day. The Foundry post narrows the statement. It says the harness in Foundry Agent Service is in public preview. The same Foundry post says general availability, this week, applies to long-running resilience for hosted agents. Available, public preview, and general availability stay as three separate phrases.
[1][2]
The list of products sits on the Microsoft blog. The harness powers Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI, VS Code, and more. The page defines a harness as the layer around the model: instructions, context, an execution environment, and controls. It says that layer supports open standards, naming MCP, agent skills, and agent plugins. It also says the harness covers more than 20 frontier models, plus a key supplied by the customer. That sentence does not name the models.
[1][2]The runtime was recently ported to Rust. The page says the port brought reductions in session overhead and in time to first token. It does not say how large those reductions are. The scale claim is the company's own: millions of users, and hundreds of real-world use cases. The SDK can be downloaded, or the harness can run in the cloud, managed by Foundry. The same post says the platform stays open, and that another harness or framework can still be brought in. The blog therefore does not present this harness as the only entrance. Microsoft runs the agent loop and the sandbox. Instructions, skills, tools, and the model choice come from the side that is building the agent.
[1][2]The Foundry post splits the public preview into four building blocks. Reusable skills package instructions, scripts, and supporting resources into procedures. Connected tools cover supported Foundry tools and Model Context Protocol tools, with authentication and approvals configured explicitly. Execution and delegation let agents work with files and code, and let them hand bounded tasks to built-in subagents while a main agent coordinates the result. Agent definitions can sit in source control beside application code. Those four blocks describe a product structure. They are not a scorecard.
[1][2]Long-running recovery is a second sentence, separate from the preview. Hosted agents let a customer bring a framework, an agent loop, middleware, dependencies, and runtime behavior, while Foundry manages hosting. Long-running resilience for those hosted agents is generally available this week. The page ties it to work that takes hours or days. Foundry detects an interruption and restarts the work with the original input. Saving progress and intermediate results is left to the application. If the agent also updates another system, the application has to keep a restart from repeating that action. The page splits the recovery foundation from the decision of what to save.
[1][2]A customer sentence stays a quotation. The Foundry page quotes Chris Evans, chief technology officer at Confluence Technologies, that a solution was running in less than a week. The page does not attach a sample size, a timing method, or a comparison. Less than a week remains inside the quotation. It does not become a measurement.
[1][2]The two posts do not collapse into one launch status. The Microsoft blog says the harness is available through Foundry on the day of the post. The Foundry post says public preview, and it reserves general availability for long-running resilience. Neither post states the size of the Rust reduction, the names behind the count of more than 20 models, the day the preview ends, or a third-party measurement behind the millions of users. The picture with this column is the diagram on the Microsoft blog: a hexagon on a dark field, with dotted lines spreading outward.
[1][2]