How to think about Pylon AI agents - Our philosophy

Within Pylon there are two fundamental concepts of AI agent use cases: deflection mode and runbooks. Both modes have their strengths and are particularly good for specific situations.

The following flow chart gives an overview of the order of events when an AI agent gets assigned:

Important notes:

• Before the agent tries to answer the question and deflect the request, the agent checks all runbook scenarios (that are assigned to the agent) to see if a runbook is applicable. If that is the case, the runbook gets executed.

• When an agent is executing a runbook, it does not have access to your training data aka your docs, websites, kb, etc.

• All steps within a runbook are sandboxed, so unless you define a variable that carries a certain value through the subsequent runbook steps, it doesn't have context on the information gathered in the previous steps.

• There is no learning curve to your ai agents, it comes down to

  • the quality and depthness of your knowledge base

  • the scenario description in your runbooks

  • the prompting within your runbooks

Definition of a Runbook:

• Order of steps that the agent follows for a very specific scenario (think of it as a very specific SOP that the ai agent follows, similar to a human agent)

• Only use for very specific use cases where actual actions are necessary (i.e. asking for information, making an api call, transferring the ticket to another queue)

• How do I structure my Runbook?

Definition of Deflection mode:

• Access to all your knowledge at any given time (contigent on your KB settings)

• Tries to answer based on the training data (i.e. the better your KB, the higher the deflection rate!!)

• Equivalent to L1 support questions that are usually answered through 1 or two messages without a lot of back and forth

• Start with this in QA mode and gradually add complexity through runbooks or splitting up agents if needed.

To summarize