Microsoft Foundry adds Copilot harness, browser automation and resilience tools

A managed loop for more capable agents

The Microsoft Copilot harness entered public preview on October 28 through Foundry Agent Service. According to Microsoft, it gives prompt agents a managed loop and sandbox for tasks that require more than a single model response, including reading files, running code, using supported tools and applying reusable skills.

Developers still select a supported model, while the harness manages the loop that turns the model’s next step into work within the configured tools and permissions. Source-controlled agent definitions and reusable skills can keep repeatable task instructions alongside an application.

Microsoft says the harness is selected when creating an agent, and switching an existing agent afterward is not supported yet. Foundry manages the sandbox’s CPU, memory and idle settings. Existing agents are unchanged, so the preview is aimed at teams that want to evaluate a managed execution loop without rebuilding around a new model.

Resilience and browser workflows reach production status

Long-running resilience in hosted agents also reached general availability on October 28. After an interruption, Foundry can restart work with its original input, but Microsoft says applications must save the progress needed to resume. That can include intermediate results, completed steps and identifiers that prevent an external action from being repeated.

The guidance is especially relevant for workflows with side effects. Microsoft gives the example of an agent that gathers documents, creates a report and sends it for review. If the task stops after the report is saved, the application should reuse that report and check whether the review request already exists, rather than creating duplicate work. Microsoft recommends testing interruptions at each side-effect boundary and pairing checkpointing with idempotency.

Browser Automation became generally available on October 19. It uses a separately provisioned Playwright Workspace and connects through a Toolbox remote MCP endpoint. The tool supports prompt and hosted agents, public and private websites and private-network connectivity configured with the workspace. Microsoft recommends least-privileged browser identities, treating browser content as untrusted input and requiring human approval for consequential actions such as submitting a transaction or changing access.

Models, decisions and upgrade considerations

Microsoft also lists GPT-6.1 Terra as generally available in Foundry from October 26. Through the Responses API, the model accepts text and images for visual reasoning and screen-based agent tasks. The October roundup also identifies the Decisions API as generally available, with GPT-6 Luna available through the API for predicates, fixed choices and rubric scores.

The company says MAI Image 2.6 is in preview in six Foundry regions, while an audio launch introduces streaming transcription and multilingual voice models. Foundry Local ships SDK and runtime 2.2.1, alongside a separate CLI preview.

For developers, Microsoft warns that October’s SDK and CLI changes include preview API renames, changed evaluation commands and coupled hosted-runtime dependencies. The update advises checking changed CUDA dependencies before upgrading GPU applications. The combined release gives teams more managed building blocks, but the source also makes clear that approvals, checkpoints, safe handling of repeated work and deliberate upgrades remain application responsibilities.