Real time ELIF+ visual introspection of cohorts of agents

[shiloh92]

🚀 Feature

In order to combat blackboxing and ignorance of the operations of autogpt style agents and other directed components, it is highly desirable to have visual bird's eye view into real time activity and planning tasks.

Motivation

Improve the utility and feedback loop of agents ensuring robust controls for the end user. The more granular the control, the better - but up to a point. A visual view that provides insights is a powerful aid for most users.

Additional context

This tweet shows a very simplified view of what an agent created. This could be a "ELIF view" : https://twitter.com/i/status/1647741551534125057

Another view could be more like a traditional mind map with just a few buttons to allow "directing" or "supervising" the agents: https://dribbble.com/shots/16089933-Mind-Palace-App-Map

Buttons suggested for robust layer of control for end user:

• TURTLE MODE, SLOW MODE, MODERATE SPEED, FULL SPEED ensures human can keep up with numerous agent cohorts
• EDIT BUTTON next to each bubble item
• CANCEL BUTTON cancels the branch or specific item
• FULL STOP BUTTON with a fire alarm bell icon, self explanatory

[gorarakelyan]

Hey @shiloh92!

This sounds great! We are currently focused on shipping Aim v4, which provides an interface to build modular metadata applications - defining metadata types and building custom dashboards with python. Afterwards we are going to focus on adding capabilities for tracing and managing AI systems based on Aim v4 groundwork.

Meanwhile, would you like to try out an early alpha version?

Also, we just started tuning Aim for AI agents tracing. I would be supper happy to jump in a quick call to learn more about your use cases and share our vision/direction.