Microsoft has introduced the Microsoft Copilot harness in Copilot Studio, designed to help agents tackle complex, multistep business tasks. The harness acts as an engine that ensures agents complete tasks correctly, even when they encounter obstacles. Unlike the Standard harness, which is suited for quick, conversational agents, the Microsoft Copilot harness is built for longer, more intricate work that spans multiple steps and involves large datasets.
The harness operates through a continuous loop, where the model receives a task, decides on a step, executes it, checks the result, and determines the next action. This process repeats until the task is completed or a set limit is reached. The design includes a secure container with real tools, such as a command line, code execution, and a file system, allowing agents to handle large data efficiently without squeezing it into a single prompt.
Security is a key feature, as the agent does not have direct access to systems or credentials. Copilot Studio manages all access and policy, ensuring organizational control while giving the agent the freedom to think and work. The team measured the harness's performance against the Standard harness, finding that agents using the Microsoft Copilot harness passed 78% of tasks with an OpenAI model and 89% with an Anthropic model, representing gains of up to 22 percentage points.
Additionally, the harness improves tool reliability without requiring maker tuning. For example, the Excel Online connector's success rate increased from 85.6% to 96.3% through continuous improvements. Microsoft plans to announce further innovations at Microsoft Ignite 2026.