Microsoft is extending a shared foundational harness across Copilot experiences, bringing powerful advances in agentic reasoning, tool use, and agent orchestration to products across Microsoft. Microsoft Copilot Studio makers are already building with this technology through what was previously called the GitHub Copilot harness. Now, to reflect its broader role across Microsoft, we’re updating the name to the Microsoft Copilot harness.
For Copilot Studio makers, while the harness name is changing, the innovation behind it isn’t slowing down. The Microsoft Copilot harness is already powering some of our most exciting advances in agent capabilities, combining sophisticated reasoning, tool use, and orchestration to help makers build agents designed to tackle complex challenges and take meaningful action.
Your existing agents will continue working as they do today, with no changes to capabilities or billing. Over the coming weeks, you’ll begin seeing the updated terminology throughout the product.
Microsoft Copilot harness enables better agents, backed by data
A capable agent needs to do more than generate a good response. It needs to navigate complex tasks, use tools effectively, and make meaningful progress toward a goal. In internal evaluations comparing the Microsoft Copilot harness with the Standard harness, we’re seeing compelling results:
- Better results across different models: In our evaluations, changing models had little impact on pass rates, but changing harnesses made a substantial difference. Compared to the Standard harness, the Microsoft Copilot harness improved evaluation pass rates on the same tasks by 10.8 percentage points using an OpenAI model and 22.1 percentage points with an Anthropic model.1
- Greater resilience when things go wrong: When a tool call failed, agents using the Microsoft Copilot harness continued trying to solve the problem in 84–91% of evaluated cases, compared with fewer than 7% using the Standard harness. And in many cases, persistence paid off: these agents were more than 7 times more likely to successfully retry the failed tool later in the conversation, compared to the agents on the Standard harness.2
- More work, more efficiently: Even while making nearly twice as many model calls, the Microsoft Copilot harness used more efficient caching to reduce the net cost of processing input by 15–21%, helping agents tackle complex tasks more cost-effectively.3
These findings demonstrate how a harness can impact agent performance. By helping agents use reasoning models more effectively, recover from failures, and make better use of model resources, the Microsoft Copilot harness is helping organizations unlock more value from the models they already use.
And we’re building on that foundation to help makers tackle even bigger challenges. Recently, we announced new capabilities for building persistent, proactive agents in the flow of work, including agent accounts, business process memory, and triggers in the Microsoft Copilot harness.
Together, these capabilities open up new possibilities for delegating ongoing business responsibilities to agents, while keeping people in control of important decisions. For Copilot Studio makers, it’s another way the Microsoft Copilot harness can help turn ambitious ideas into practical business solutions.
Real results, real-world impact
The possibilities are exciting, but what’s already happening in practice is just as compelling. At Microsoft, we’re seeing firsthand what agents powered by the Microsoft Copilot harness can accomplish. Microsoft Customer Service and Support Business Applications built a dynamic resolution agent that uses reasoning to analyze support cases, draws on existing knowledge and tools. In August 2026, the agent achieved a fully autonomous resolution rate—defined as resolving cases without human intervention—of over 35%.
In this video, the team shares how they built the agent, the breakthroughs they experienced, and lessons that could help other organizations rethink their approach to agent-powered business processes:
This is more than a hypothetical scenario. It’s a look at how agents are being put to work in a real support environment—an environment where accuracy, reliability, and customer experience matter every day.
Keep building with Copilot Studio
The Microsoft Copilot harness may have a new name, but our focus remains the same: helping makers build increasingly capable, reliable agents that deliver real business value.
With continued investments in harness quality, extensibility, and new agent capabilities, we’re excited to keep bringing the latest innovations to Copilot Studio.
And as the examples above show, there’s no need to wait to get started.
Take your agents further
Grow your Copilot Studio skills with practical, hands-on training in Agent Academy.
1 Source: Microsoft internal analysis of agent evaluation data, October 2026. Results are based on evaluations comparing the Microsoft Copilot harness with the Standard harness across equivalent tasks and models. Figures reflect results from successful evaluation runs and may vary by day, build, model, and configuration.
2 Source: Microsoft internal analysis of production telemetry, October 2026. Findings compare agent behavior following failed tool calls across the Microsoft Copilot harness and Standard harness, including continued tool use and subsequent successful retries. Results do not necessarily indicate successful completion of the overall task. Results may vary by scenario, model, and channel.
3 Source: Microsoft internal analysis of agent evaluation data, October 2026. Results are based on evaluations comparing the Microsoft Copilot harness with the Standard harness across equivalent tasks and models. Figures reflect results from successful evaluation runs and may vary by day, build, model, and configuration.
