Xueyang Feng, Chaozhuo Li, Xiaohe Bo, Jiakai Tang, Chenxi Li, Qiwei Ye, Xu Chen
POrchestra introduces a dynamic framework where sub-agents can request adjustments to their execution harnesses during runtime, enabling adaptation to changing environments. It uses hindsight feedback from these requests to train the orchestrator via On-Policy Orchestration Distillation, improving policy learning. The approach boosts task success rates and offers better cost-performance balance across diverse long-horizon tasks.
Dynamic task execution in real-world settings requires agents that can adapt mid-process. POrchestra enables this adaptability while improving training efficiency through targeted feedback, making large-scale agent systems more robust and practical. This advances the deployment of autonomous agents in complex domains like research, software development, and digital assistance.
Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.