Sub-agents can now halt mid-task and ask for a new harness: 55.5% on GAIA2

A sub-agent is no longer locked to the harness it was created with: it can pause at any logical point, ask the orchestrator to adjust its task scope, working context, tool set, or execution budget, and then resume with its prior execution history intact. Earlier re-orchestration methods had to wait until the current sub-agent submitted its result or timed out before creating the next one with a new harness, which delayed adaptation and discarded the sub-agent's ongoing execution state. On the more dynamic GAIA2, POrchestra reaches 55.5% with DeepSeek-V4-Flash, a 24.6% relative improvement over the strongest baseline ReAct at 44.5%; AOrchestra, by contrast, scores 76.4% on GAIA against ReAct's 69.1% but only 39.8% on GAIA2 against ReAct's 46.9%.

The structured feedback sub-agents send back is also used to train the orchestrator itself, turning obstacles met during execution into improvement signals for earlier orchestration decisions. Prior supervision for the orchestrator was either a sparse reward available only after the task ended or a local reward that judged output compliance or decision quality, offering little attribution to individual orchestration decisions. Training Qwen3.5-9B orchestrators, OPOD brings POrchestra to 77.9% on GAIA and 49.2% on GAIA2, 5.5 and 5.4 percentage points above its SFT initialization; removing process feedback drops GAIA2's Adaptability and Time tasks by 7.8 and 7.9 points, and removing outcome supervision lowers GAIA and GAIA2 averages by 2.1 and 3.1 points.

The extra cost of proactive communication is offset by savings in orchestration and execution, so total inference cost comes out lower. Earlier orchestration systems carried high context overhead in environments like GAIA2 with its large tool set and dense application feedback, and wasted inference when harnesses did not match. 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.

The mechanism is evaluated on three long-horizon benchmarks, GAIA, GAIA2, and SWE-Bench Verified, covering deep research, dynamic digital assistance, and software engineering; GAIA uses the 165-task 2023 validation set, GAIA2 uses 128 mini scenarios, and SWE-Bench uses a fixed subset of 100 issues. Inference comparisons use Gemini-3-Flash as orchestrator with Gemini-3-Flash, DeepSeek-V3.2, and DeepSeek-V4-Flash as execution backbones; training comparisons use Qwen3.5-9B orchestrators with DeepSeek-V4-Flash as executor. It is aimed at researchers and engineers working on multi-agent orchestration and orchestrator policy optimization.

Only the orchestrator is optimized, while sub-agent model parameters and communication rules stay fixed; the authors list learning a general proactive-communication capability that transfers across models, tasks, and dynamically generated sub-agents as future work. In the complex interaction tasks studied, non-concurrent orchestration baselines outperform the concurrent baseline, which motivates keeping a single active sub-agent at a time; concurrent execution still offers potential efficiency gains for independent subtasks, and coordinating feedback and revision requests across concurrent sub-agents is left open. The authors also note that agents able to revise goals, tools, context, or budgets introduce risks when deployed with access to external systems, including unintended actions, privacy exposure, and misuse of granted permissions, so proactive revision requests should remain subject to explicit authorization boundaries with human oversight for consequential actions.