Dynamic runtime alteration of LLM configs (temperature, reasoning effort, etc)

Can we have more control over how a supervisor agent assembles its message to its collaborators? We'd like to be able to make this more consistent so there is less variance in the user experience and results can be more easily replicated.

I've attached a couple of images showing how the user input into a supervisor agent can be static, but the transformed "message" string the agent then provides to its collaborator can vary.

A pre-invoke plug-in could be used for each of the collaborator agents to modify the input message, but this would result in a deterministic restructuring of the message parameter, which is great for specific user prompt injections (e.g. always ending with "Include source URLs."), but not for ensuring that messages from a wide array of domains get transformed with less variance. You can imagine the infinite number of input messages someone might provide, so having if/else statements to augment the message based on certain keywords is not scalable.

Similarly, the behavior instructions and agent descriptions can have a great influence on how the message parameter is assembled, but being able to decrease the overall variance of the resulting message is the primary goal, and LLMs evaluating descriptions/instructions is implicitly inconsistent.

Ideally, the user could be given the ability to alter the LLM's configurations (reasoning effort, temperature, top p, etc) dynamically such that when it gets to the "message creation" step, the temperature could be turned all the way down so that there is less variance in the resulting message. Combined with specific, focused agent descriptions and behavior instructions, we would likely see much less variance in how this parameter is constructed, and therefore be able to provide a more consistent experience for end users.