Microsoft tells that its new Copilot harness adds 10.8 to 22.1 points

Microsoft tells about a new harness (control software) for agents in Copilot Studio. It is one model in a safe container, with command line and tools for documents. Microsoft tells that its success rate on its tests increases 10.8 points on GPT 5.5 Chat and 22.1 points on Opus 4.8. The base is the Standard harness. No other group gave a test.

How it works

The Standard harness is made for chat. It makes some isolated model calls for each turn. It gives control to the user when a step must get input.

The new harness is one loop. It can put large data on a disk, and the model works on it step by step.

The numbers

More cache use cuts the token cost 15% to 21%. After a tool is defective, the new agents continue 84% to 91% of the time. The Standard harness continues less than 7% of the time.

What is not known

On a benchmark for company tasks, Microsoft tells that it had more points than a competitor that it does not name. The paper is by Microsoft, and no other group made a test of the results.

Insight: Microsoft needs daily use, not only paid seats

Microsoft has more than 30 million paid seats of Copilot, but only 18% of workers prefer it when ChatGPT is also open to them. This paper shows long tasks with more tool use as a method to get more use. It reads as a move from paid seats to daily use, and the harness is the tool for it. For leaders: measure use, not seats.