The Three Laws of Agentic AI Systems
OpenAI recently came out with a paper titled "Practices for Governing Agentic AI Systems" introduced in the post at https://openai.com/research/practices-for-governing-agentic-ai-systems
It feels like it is meant to push forward the conversation on Agents and get many of the issues down in black and white. It is not a hard paper to read so I suggest you give it a shot if at all interested.
A couple of interesting points include the framework for describing AI agents and their "agenticness" including:
- Goal complexity
- Environmental complexity
- Independent execution
And their "Practices for Keeping Agentic AI Systems Safe and Accountable" which include:
- Evaluating Suitability for the Task
- Constraining the Action-Space and Requiring Approval
- Setting Agents’ Default Behaviors
- Legibility of Agent Activity (note they don't say explainability)
- Automatic Monitoring
- Interruptibility and Maintaining Control
All in all a good overview that highlights key issues. There will be a lot of research in this area and the blog post mentions a new grant program for studying agentic AI systems practices.