Computer-use AI agents struggle with multi-step workflows like email and customer support…

TL;DR - Microsoft Research announces Echoverse, a training approach for computer-use AI agents that emphasizes realistic, evolving environments over simply scaling up the number of training tasks. It matters because current computer-use agents reportedly break down on long, multi-step workflows like handling email or customer support.

• Framed problem: computer-use agents underperform on multi-step, real-world desktop/web workflows (email, customer support).
• Proposed approach: train in realistic environments rather than just adding more training tasks.
• Co-evolution angle: tasks, tests, and environments evolve alongside the agent so it keeps improving.
• Content is thin — a promotional post with a link and video, so no benchmarks, metrics, or architecture details are provided.