Sub-agents can now ask to be reconfigured mid-task, lifting long-horizon success by 9.1%
Synopsis
POrchestra lets sub-agents proactively send structured requests to the orchestrator during execution, so the orchestrator can revise their subgoal, context, tool set, or execution budget on the spot instead of waiting for the sub-agent to finish or fail. Earlier re-orchestration methods fixed each sub-agent's configuration until its lifecycle ended, delaying adaptation to dynamic environments. Across GAIA, GAIA2, and SWE-Bench Verified, task success improves by an average of 9.1% over the strongest baseline, and the companion OPOD turns that feedback into process supervision, raising small-orchestrator success by an average of 10.0% over training baselines.
Interpretation
A sub-agent can pause at any logical point during execution, ask the orchestrator to revise its execution configuration, and then resume with its accumulated execution history intact. Earlier re-orchestration methods had to wait until the current sub-agent submitted its result or timed out before creating a new sub-agent with a new configuration, which delayed adaptation and discarded the ongoing execution state. On GAIA2 with DeepSeek-V4-Flash, POrchestra reaches 55.5% success, a 24.6% relative improvement over the strongest baseline ReAct at 44.5%, and it also attains the highest success on GAIA and SWE-Bench.
Proactive feedback is converted into targeted improvement signals for preceding orchestration decisions and combined with task-level outcome rewards to optimize the orchestration policy. Earlier orchestrator optimization either relied on sparse, delayed task-level rewards or used format-based and rubric-based local rewards that lacked decision attribution grounded in the obstacles actually encountered. With a Qwen3.5-9B orchestrator, OPOD brings POrchestra to 77.9% on GAIA and 49.2% on GAIA2, improving over its SFT initialization by 5.5 and 5.4 percentage points; removing process feedback drops GAIA2's Adaptability and Time subsets by 7.8 and 7.9 points.
The overhead of proactive communication is offset by lower routine orchestration and sub-agent execution costs, leaving total inference cost below the compared systems. Earlier orchestration systems carried higher context and execution overhead in environments like GAIA2, with its large tool catalog and extensive application feedback. On GAIA2 with DeepSeek-V4-Flash, POrchestra exceeds ReAct by 11.0 percentage points at 36% lower per-task API cost and Orchestra-o1 by 11.7 points at 31% lower cost.
Perspective
The reported scope is long-horizon agent orchestration: deep research on GAIA, dynamic digital assistance on GAIA2, and software engineering on SWE-Bench Verified, with Gemini-3-Flash, DeepSeek-V3.2, and DeepSeek-V4-Flash as execution backbones and Qwen3.5-9B as the trained orchestrator. It suits readers following multi-agent orchestration, agent training, and inference cost. Only the orchestrator is optimized; sub-agent model parameters and communication rules stay fixed, and the authors list learning a general proactive-communication capability that transfers across models and tasks as future work.
The authors note that only the orchestrator is optimized while sub-agent model parameters and communication rules remain fixed, and that whether a general proactive-communication capability transfers across models and tasks is still open. On concurrency, non-concurrent orchestration baselines outperformed the concurrent baseline on the complex interaction tasks studied, so a single active sub-agent is kept at a time, leaving coordination of feedback and revision requests across concurrent sub-agents unresolved. Agents that revise goals, tools, context, or budgets can introduce unintended actions, privacy exposure, and misuse of granted permissions when connected to external systems, so the authors recommend explicit authorization boundaries and human oversight.