You can train a separate agent to supervise and correct LLM systems in real-time, catching goal drift before it wastes resources—and this steering intelligence transfers to new tasks and different executor models without retraining.
This paper introduces a steering agent that monitors LLM-based systems during execution and intervenes when they drift from their goals. The steering agent learns when to intervene (fast path) and how to intervene (slow path) using reinforcement learning, improving performance and efficiency for both single and multi-agent systems while generalizing to new tasks.