The DRAFT framework decouples the safety judgment process into an Extractor that creates a compact latent representation of long agent trajectories and a Reasoner that attends to both the draft and full context. This addresses the core difficulty of sparse risk-critical evidence in noisy interactions, achieving 91.18% average accuracy compared to 63.27% for baseline LoRA fine-tuning on ASSEBench and R-Judge (Wang, 2026).

Original coverage of the abstract fails to connect this to broader patterns in agent deployment, such as the increasing use of computer control agents by Anthropic and OpenAI, where long-horizon tasks amplify the supervision challenge. Related work like ReAct (Yao et al., 2022) demonstrated the power of interleaved reasoning but left safety auditing underdeveloped. DRAFT builds on latent variable models to enable differentiable credit assignment where binary labels fall short.

Synthesizing with Constitutional AI (Bai et al., 2022), which uses AI feedback for harmlessness, reveals DRAFT's novelty in performing evidence aggregation entirely in continuous latent space rather than discrete summaries. This synergy points to robust paths for oversight as agents gain real-world autonomy, identifying a missed opportunity in prior research for end-to-end trainable safety systems tailored to trajectory-based supervision.