Challenges and open questions
In this section, we will address several open questions about both agents and the capabilities of LLMs. Despite advances in the field, several points remain to be resolved for the safe use of AI agents.
Challenges in human-agent communication
Once they are deployed in the real world, agents can perform actions that lead to problematic failures. For example, a shopping agent might spend money unexpectedly or inadvertently leak sensitive information. Coding agents might execute or produce viruses, delete important files, or push repositories into production that are full of bugs. Communication with the user is key to avoiding such problems. The use of agents should be based on two key principles: transparency and control. Indeed, there must be an alignment between the user’s goals and the agent’s behavior; the user must then be able to control the process and have access to its progress. Communication between humans and agents allows...
