AgentBrew Distills Teacher Knowledge for Weaker LLM Agents

Deploying Large Language Model (LLM) agents often necessitates using a compact, less powerful student model at test time, even if a more capable teacher model was available during development. The AgentBrew framework introduces a novel approach called "knowledge brewing," which distills the interactive experience of a strong teacher agent into an external, persistent memory for a weaker student agent. This process is notable because it requires no weight updates, expert demonstrations, ground-truth labels, or access to the teacher model during deployment. AgentBrew tackles two primary challenges: the sparse, binary feedback typically provided by environments, and the need to calibrate teacher-generated knowledge to be concretely executable by a significantly weaker student. It achieves this through two core components: the Ralph Loop, a failure-triggered teacher mechanism that transforms student failures into environment-validated notes, and student-aware synthesis, which tailors the teacher's knowledge to the student's operational granularity. Extensive evaluations across coding, math, and tool-use tasks demonstrate that this asymmetric, training-free distillation paradigm produces highly capable and deployable LLM agents.