Microsoft Copilot Harness — A Shared Execution Foundation Across Copilot Studio and Microsoft Foundry.
Back in August, I wrote about the different harnesses available in Microsoft Copilot Studio: Standard, Copilot Chat and GitHub Copilot. I also shared Microsoft's Agent Academy learning paths for getting hands-on with them.
Microsoft's October 28 announcement brings an interesting update to that earlier discussion.
The GitHub Copilot harness in Copilot Studio is now called the Microsoft Copilot harness.
But there is more to this announcement than a name change.
What Has Changed?
Microsoft is extending a shared harness foundation across its Copilot experiences, including Microsoft Foundry.
In Copilot Studio, the change is primarily about naming. Existing agents continue working without changes to their capabilities or billing, and the updated terminology will appear over the coming weeks.
In Microsoft Foundry, however, Microsoft has introduced the Copilot harness in Foundry Agent Service in public preview.
It provides a managed execution environment for more complex agents, bringing together reusable skills, connected tools, file and code execution, subagent delegation and working-context management. Developers can also manage agent definitions alongside their application code.
Separately, Microsoft announced the general availability of long-running resilience for hosted agents, helping workflows recover from interruptions. Applications still need to manage their own checkpoints and avoid duplicate actions when restarting.
Understanding the Harness Evolution: Renaming in Copilot Studio, Shared Capabilities, and Expansion into Microsoft Foundry.
Why I Find This Interesting
When I first explored harnesses, the discussion was mainly about choosing the right approach for building an agent.
The Standard harness was useful for more structured, predictable workflows, while the GitHub Copilot harness supported more flexible, reasoning-intensive tasks.
Now Microsoft is extending that execution foundation beyond Copilot Studio into Foundry.
Microsoft also shared some interesting results from its internal evaluations:
- Changing harnesses made a greater difference to task pass rates than changing models in the scenarios Microsoft evaluated. Compared with the Standard harness, Microsoft reported improvements of 10.8 percentage points with an OpenAI model and 22.1 percentage points with an Anthropic model.
- After a failed tool call, agents using the Microsoft Copilot harness continued attempting the task in 84–91% of evaluated cases, compared with fewer than 7% using the Standard harness.
These are Microsoft's internal evaluation and telemetry results, not independent benchmarks. The results may vary by model, configuration and scenario, and continued attempts do not necessarily mean the overall task was completed successfully.
Still, they reinforce an important learning: a more capable AI model alone doesn't guarantee a more reliable agent. The execution system around it matters too.
A Note on Preview Availability and Cost
The Microsoft Copilot harness in Foundry Agent Service is currently in public preview.
During this preview, the harness must be selected when creating an agent. Existing agents remain unchanged, and switching the harness after creation is not yet supported.
Harness runtime usage is free until December 1, 2026, after which consumption-based runtime billing begins. However, model inference, separately billed tools and connected services may still incur charges during the free-runtime period.
This is separate from Copilot Studio's existing billing arrangements, which Microsoft says are unaffected by the rename.
Connecting This to Agent Ownership
Microsoft's Copilot Studio announcement also points to persistent, proactive agents using capabilities such as agent accounts, business process memory and triggers.
As agents take on ongoing business responsibilities, their execution capabilities must be accompanied by clear accountability.
This connects to another topic I explored recently: Who Owns an AI Agent After It Goes Live? Building the Enterprise Agent Operating Model.
An agent that can continue working, retain context and execute actions still needs a named owner, appropriate permissions, monitoring and a way to intervene when something goes wrong.
My Perspective
For me, this announcement connects the learning from August with where Microsoft's agent platforms are heading.
We started by understanding different harness types and how to build agents with them. Now we're seeing Microsoft extend a shared execution foundation across its broader AI ecosystem.
For organizations adopting agents, the starting point should remain the business scenario, required controls, complexity and cost—not simply which harness is newest.
The model provides intelligence. The harness helps turn that intelligence into controlled execution..


No comments:
Post a Comment