←Home

    Agents at Work

    A guided tour of current research on AI agents: how they are measured on real work, the harnesses, memory and skills built around models, how they learn from experience and improve themselves, where self-improvement goes wrong, and what it takes to act on real devices and real tasks.

    Background: Memory for AI agents

    In order

    1 stories
    手術醫師, 藍光
    手術醫師, 藍光

    Photo: Navy Medicine

    Agency · Agent benchmarks

    AI agents fully finish at most a quarter of expert-built professional tasks

    A benchmark built from real projects across 55 kinds of digital work finds the best agent setup fully passes 24% of public tasks, and agents pass almost none of the hardest.

    arXiv preprint · 3 June 2026