Proactive Agentic Orchestration
Xueyang Feng, Xu Chen and colleagues (Gaoling School of AI, Renmin University of China) present POrchestra, an orchestration framework in which sub-agents can ask the orchestrator to change their goal or harness mid-task, and OPOD, a training method that turns those requests into process supervision for the orchestrator.
Ask this paper
Fixed sub-agents break in dynamic tasks. Existing orchestrators set a sub-agent goal and harness at creation time. When task requirements change, AutoAgents scores 0% on the adaptability tasks with all three backbones tested.
Proactive revision requests. A sub-agent can message the orchestrator at any step to request an adjustment while keeping its local execution state. Proactive feedback recovers 92% of disrupted tasks with a mean delay of 1.67 steps.
OPOD. On-Policy Orchestration Distillation attributes each revision request to the orchestration decision that caused it and uses that as a process signal alongside the final reward.
Results. Across deep research, dynamic digital assistance and software engineering, POrchestra improves task success by 9.1% on average over the strongest baseline. On GAIA2 with DeepSeek-V4-Flash it reaches 55.5% against 44.5% for ReAct, at 36% lower per-task API cost.
Small orchestrators. With OPOD, a trained small orchestrator reaches 77.9% on GAIA and 49.2% on GAIA2, 5.5 and 5.4 points above its SFT initialization.
Abstract
Agentic orchestration delegates tasks to sub-agents with independent contexts, providing a key path to extending the effective execution horizon of large language model agents. However, existing orchestration methods face two limitations: in execution, current methods fix sub-agents' goals and harnesses at creation time, limiting their adaptability to dynamic real-world environments; in optimization, sparse and delayed feedback constrains effective learning of the orchestration policy. To address these limitations, we introduce POrchestra, an on-the-fly proactive agentic orchestration framework. It allows sub-agents to proactively communicate with the orchestrator at any logical point to request harness adjustments while maintaining continuity of local execution. This mechanism not only enables sub-agents to adapt seamlessly to dynamic environments but also naturally inspires On-Policy Orchestration Distillation (OPOD) based on hindsight feedback. OPOD attributes sub-agents' proactive revision requests back to preceding orchestration decisions to derive targeted improvement signals, and leverages these signals as process supervision alongside the final reward to optimize the orchestration policy. We evaluate POrchestra on three long-horizon agent benchmarks spanning deep research, dynamic digital assistance, and software engineering. POrchestra improves task success by an average of 9.1% relative to the strongest baseline; combined with OPOD, it improves small-orchestrator success by an average of 10.0% relative to existing training baselines. POrchestra also achieves a more favorable cost-performance trade-off. Our code is publicly available at https://github.com/XueyangFeng/POrchestra.