Microsoft has made the agent-execution harness used across its Copilot products available to developers through Microsoft Foundry Agent Service. The October 28 public preview packages the orchestration around a model — context management, tool use, sandboxed execution, recovery and long-running task control — as a managed service.
The company says the same harness now underpins Copilot Code, Copilot Studio, the GitHub Copilot app, the command-line interface and Visual Studio Code. Microsoft describes support for more than 20 frontier models and bring-your-own-key configurations, so the launch is about standardizing the agent loop rather than locking developers to a single model family.
The accompanying Foundry announcement also makes long-running resilience for hosted agents generally available. That capability is intended to preserve state and continue work through interruptions, which matters when an agent performs multi-step tasks instead of returning a single response.
Microsoft says it ported the harness runtime to Rust and reports lower session overhead and faster first-token delivery. Separate Copilot Studio material cites internal evaluation gains in task completion and recovery from failed tool calls. Those measurements are vendor-run and may not generalize across workloads, models or configurations.
For developers, the practical change is that Microsoft can operate the agent loop and sandbox while teams provide instructions, skills, tools and model choices. The SDK is downloadable, and the harness can also run as a managed Foundry service.