Скачать или смотреть THOUGHT LEADERSHIP: Building Trustworthy Computer Use Agents

This talk explores the paradigm shift from traditional AI assistants to autonomous computer use agents – AI systems that can independently navigate digital environments, control interfaces, and execute complex workflows on behalf of users.

We’ll examine how recent breakthroughs in multimodal large language models have enabled agents like OpenAI’s Operator, Anthropic’s Computer Use, and emerging open-source frameworks to interact with graphical user interfaces as humans do, using visual perception and natural language understanding.

The talk will trace the evolution from foundational AI agent concepts—planning, tool use, and reflection—to today’s sophisticated systems capable of multi-step reasoning, error recovery, and cross-application automation.

We’ll also cover the critical technical challenges that make computer use agents both powerful and risky, including:

Agent hijacking through prompt injection
Value misalignment in autonomous execution
Privacy concerns across different contexts, and
The fundamental tension between capability and control