Large language model agents often rely on hand-crafted workflows to break down complex tasks. The authors of a new arXiv preprint argue that this manual approach is time-consuming and struggles to adapt to changing conditions.
To address this, they introduce CB-Orchestrator, a framework that frames workflow optimization as a contextual bandit problem. This allows the agent to select and adjust workflows based on context, rather than following a fixed procedure.
The work is presented as a way to make agentic workflows more flexible and less dependent on human intervention. The abstract does not include experimental results or implementation details, so the claims are limited to the proposed approach.