Salesforce EVP of AI @justjayesh says we may be surprisingly close to a full self-improvement loop for AI agents without touching the model weights at all: "Most of the conversation I see about RL, recursive self-improvement, RSI, is on the model layer, and it needs to be thought through more holistically." "That success criteria can be met by having better context. It could be resolved by having a workflow that the agent has access to so it doesn't have to go cook one up on the fly. And of course the weights can also improve." "We are pretty near doing a full improvement loop using mechanisms that are modifying context by themselves, modifying prompts within the agentic system by themselves, and freezing the weights for a minute." "This Dreamforce we announced Agentforce Optimizer, which takes the very first step on an agent that's in production and just tells you where the agent is failing and why it might be failing." @salesforce @Dreamforce