SOB Life Cycle - Agent Review Process

[GonzoDMX]

I am attempting to build a prototype for the SOB and I wanted to get some input on the life cycle for this element in the swarm. I believe the long term goal is that the SOB will be self-directed, however, for prototyping I think certain aspects of the SOB's behavior need to be scripted. This is also to avoid burning through all my API credits.

I have other questions but will break them out into separate discussion posts in order to keep each discussion focused on a single element.

1. By what criteria should an agent come under review by the SOB?

For prototyping I was either thinking a time scheduled review or a counter based review cycle. For example, after an agent performs 10 actions the SOB will analyze the trend of those 10 actions and judge whether it conforms to the swarm's governance policy. Alternatively, the SOB could ask questions directly to the agent under review and determine whether its answers indicate alignment with the governance policy. These approaches could also be abstracted if there are multiple governance layers in the swarm.

If I am missing any concepts don't hesitate to direct me to academic papers or references that will fill in the gaps.

[Gottheil]

One idea just came to mind that might help in terms of limiting the operational cost of the oversight sparked by your credits comment.

The board probably only needs the high level summary of agents actions in order to govern them. I am envisioning that every outward acting agent produce a sussinct description of the actions it is about to undertake before acting, that might as a side benefit also ensure more robust performance of the agents train of thought. Those small snippets might be computationally small enought to be directly handled by the SOB(or perhaps a randomly selected member of it). If anything hits as a threat to the current consensus of the board you could trigger an investigation into the whole branch where that signal came from.

Basically: fully or sporadically check any actions executed and if anything registers as off target/threat investigate everything in contact with that node.