Microsoft publishes Copilot Studio harness evaluation for complex business tasks
Microsoft published a technical paper comparing Copilot Studio’s conversational Standard harness with the Microsoft Copilot harness for longer tasks that use code, files and repeated tool calls.
The latter runs a model-centered agent loop inside a secure container and can work iteratively with large datasets on disk.
In the reported evaluations, better caching reduced cost-weighted token consumption by 15–21% despite twice as many model calls. Pass rates rose by 10.8 and 22.1 percentage points on the two tested models.
The post also examines recovery after failed tool calls and links the full paper for agent builders.
Partner briefing
- What it is
A technical paper evaluating Microsoft Copilot harness execution in Copilot Studio.
- What’s changing
Compares the Standard harness with a container-based agent loop using command-line tools, files and iterative data processing.
- Why it matters
Reports evaluation-specific quality, caching, latency and tool-recovery measurements for complex business tasks.
- Who it affects
Copilot Studio makers and developers building agents for multistep business tasks.