Bootstrapping Agentic Behaviour

Hack Session

About the session

Most people are aware of simple LLM based workflows and of agentic architectures. The latter require you to define agents that perform specific tasks, and then allow them to interact with each other to accomplish some larger goal. However, defining these agents is a challenge, because it is not clear how to properly define their skills for every conceivable situation. I want to propose a way of bootstrapping this behaviour, using humans as trainers for agents. The idea is to begin with simpler workflows and have humans in the loop. Then one observes different patterns of human responses to questions asked by the LLMs, and then mimicking the same using agents. This allows for the definition and behaviour of agents to arise naturally.

Speaker

Download Brochure