AgentBrew is an offline training framework that learns tool-use policies from a single batch of raw interaction trajectories. The pipeline consists of three stages: 1) explore the target environment to collect unfiltered trajectories; 2) apply retrospective task inference to generate an aligned instruction for each trajectory based on its actual outcome; 3) use PMI‑based credit assignment to decompose the trajectory’s total information about the inferred instruction into additive per‑action credits. The pointwise mutual information is computed as $$\text{PMI}(a,s)=\log\frac{P(a,s)}{P(a)P(s)}$$ where $a$ denotes an action and $s$ the instruction. These credits weight the policy loss, amplifying informative actions while suppressing ineffective ones. Experiments on GitHub, Notion and PostgreSQL show that AgentBrew raises Qwen3‑32B’s average accuracy by 8.7 and score by 9.7, surpassing the gains of Qwen3‑235B and outperforming rejection‑sampling baselines. The implementation is publicly available.
Review