Unconference: Self-optimizing deep research agents

Unconference: Self-optimizing deep research agents - Jakub Zavrel - Founder and CEO, Zeta Alpha Advanced AI agents allow us to build modular iterative reasoning patterns for complex tasks on private enterprise data, such as for example a Deep Research agent. In this talk we will outline the architecture of this class of agents, and discuss how LLM-as-a-judge can be used to evaluate in the absence of gold standard benchmarking data sets. We will also show how this evaluation signal can be used for end-to-end optimization of these modular non-differentiable systems to improve performance.