'Current LLMs introduce substantial errors when editing work documents': Microsoft scientists find most AI models struggle with long-running tasks — so maybe don't trust them completely just yet

A person typing on a laptop and using a tablet. Only their upper torso, arms and hands are visible. Text superimposed on the image shows AI .

• Microsoft researchers determine that current LLMs aren't good at long-running tasks
• More interactions and less structure significantly reduce benchmark performance
• "Python is the only domain where most models are ready"

New research from a trio of Microsoft workers has uncovered a fundamental issue that could be blocking effective agentic AI -namely that most AI models can't actually reliably handle long-running workflows.