CAFE: Co-Evolving Agent and Feedback for Search Agents
August 24, 2026
CAFE uses a shared-parameter model that alternates between search-agent and critic roles to improve long-horizon search. This coupling allows the agent to learn when to request feedback and enables the critic to provide useful corrections from outcome-confounded rollouts.
HOW THIS AFFECTS YOU
●
builderThis approach helps agents recover from errors mid-trajectory rather than relying only on terminal rewards.
●
researcherYou can implement self-improving agent architectures where the critic and agent co-evolve through online RL.