Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration

Jianhao Ruan, Zhihao Xu, Yiran Peng, Fashen Ren, Zhaoyang Yu, Xinbing Liang, Jinyu Xiang, Bang Liu, Chenglin Wu, Yuyu Luo, Jiayi Zhang

TL;DR

AOrchestra introduces an orchestration-centric framework that models every agent as a dynamic four-tuple (Instruction, Context, Tools, Model) and automatically spawns tailored sub-agents for each subtask. The central orchestrator, operating with Delegate and Finish actions, decouples planning from execution, enabling on-demand specialization, context-aware task decomposition, and plug-and-play sub-agents. The orchestrator can be learned via supervised fine-tuning and improved cost-aware routing through iterative in-context learning, achieving Pareto-efficient performance-cost trade-offs. Empirically, AOrchestra delivers strong, training-free improvements across GAIA, Terminal-Bench 2.0, and SWE-Bench-Verified and demonstrates robust generalization and error recovery in long-horizon tasks. The approach is framework-agnostic and supports diverse sub-agents, offering a scalable path to robust autonomous task solving in open-world environments.

Abstract

Language agents have shown strong promise for task automation. Realizing this promise for increasingly complex, long-horizon tasks has driven the rise of a sub-agent-as-tools paradigm for multi-turn task solving. However, existing designs still lack a dynamic abstraction view of sub-agents, thereby hurting adaptability. We address this challenge with a unified, framework-agnostic agent abstraction that models any agent as a tuple Instruction, Context, Tools, Model. This tuple acts as a compositional recipe for capabilities, enabling the system to spawn specialized executors for each task on demand. Building on this abstraction, we introduce an agentic system AOrchestra, where the central orchestrator concretizes the tuple at each step: it curates task-relevant context, selects tools and models, and delegates execution via on-the-fly automatic agent creation. Such designs enable reducing human engineering efforts, and remain framework-agnostic with plug-and-play support for diverse agents as task executors. It also enables a controllable performance-cost trade-off, allowing the system to approach Pareto-efficient. Across three challenging benchmarks (GAIA, SWE-Bench, Terminal-Bench), AOrchestra achieves 16.28% relative improvement against the strongest baseline when paired with Gemini-3-Flash. The code is available at: https://github.com/FoundationAgents/AOrchestra

AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration

TL;DR

AOrchestra introduces an orchestration-centric framework that models every agent as a dynamic four-tuple (Instruction, Context, Tools, Model) and automatically spawns tailored sub-agents for each subtask. The central orchestrator, operating with Delegate and Finish actions, decouples planning from execution, enabling on-demand specialization, context-aware task decomposition, and plug-and-play sub-agents. The orchestrator can be learned via supervised fine-tuning and improved cost-aware routing through iterative in-context learning, achieving Pareto-efficient performance-cost trade-offs. Empirically, AOrchestra delivers strong, training-free improvements across GAIA, Terminal-Bench 2.0, and SWE-Bench-Verified and demonstrates robust generalization and error recovery in long-horizon tasks. The approach is framework-agnostic and supports diverse sub-agents, offering a scalable path to robust autonomous task solving in open-world environments.

Abstract

Language agents have shown strong promise for task automation. Realizing this promise for increasingly complex, long-horizon tasks has driven the rise of a sub-agent-as-tools paradigm for multi-turn task solving. However, existing designs still lack a dynamic abstraction view of sub-agents, thereby hurting adaptability. We address this challenge with a unified, framework-agnostic agent abstraction that models any agent as a tuple Instruction, Context, Tools, Model. This tuple acts as a compositional recipe for capabilities, enabling the system to spawn specialized executors for each task on demand. Building on this abstraction, we introduce an agentic system AOrchestra, where the central orchestrator concretizes the tuple at each step: it curates task-relevant context, selects tools and models, and delegates execution via on-the-fly automatic agent creation. Such designs enable reducing human engineering efforts, and remain framework-agnostic with plug-and-play support for diverse agents as task executors. It also enables a controllable performance-cost trade-off, allowing the system to approach Pareto-efficient. Across three challenging benchmarks (GAIA, SWE-Bench, Terminal-Bench), AOrchestra achieves 16.28% relative improvement against the strongest baseline when paired with Gemini-3-Flash. The code is available at: https://github.com/FoundationAgents/AOrchestra
Paper Structure (67 sections, 9 equations, 4 figures, 7 tables)

This paper contains 67 sections, 9 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multi-Agent Systems
  4. Sub-Agent as Tools
  5. Methodology
  6. Problem Formulation
  7. Sub-agent-as-tools view.
  8. AOrchestra
  9. A unified four-tuple agent abstraction.
  10. Action Space of Orchestrator.
  11. Implementation of Delegate and Finish.
  12. Advantages of AOrchestra
  13. Learnable Orchestrator
  14. Supervised fine-tuning (SFT) for task orchestration.
  15. Iterative In-context Learning for Cost-aware Orchestration.
  16. ...and 52 more sections

Figures (4)

  • Figure 1: Overall performance on three challenging agentic benchmarks (GAIA, Terminal-Bench-2, SWE-Bench-Verified) paired with Gemini-3-Flash when comparing AOrchestra against other popular agentic frameworks.
  • Figure 2: Comparison of sub-agent-as-tools approaches. (a) Sub-agents as context-isolated threads mitigate context rot but lack on-demand specialization. (b) Sub-agents as static roles provide specialized capabilities but are inflexible, leave coverage gaps, and require heavy human engineering. (c) Our Sub-agents as on-demand specialization concretizes a unified 4-tuple abstraction (Instruction, Context, Tools, Model) to enable creating tailored executors on the fly.
  • Figure 3: Overall design of our proposed agentic framework, AOrchestra, for complex, long-horizon tasks. The orchestrator solves a user task by repeatedly delegating subtasks to on-the-fly instantiated sub-agents, each defined by a unified four-tuple $(I,C,T,M)$. The orchestrator is learnable and can improve its decomposition, context routing, and capability allocation from past experience.
  • Figure 4: Pareto front curve of GAIA. We plot GAIA accuracy and average cost per task (USD, log scale). Each point corresponds to a configuration, and the dashed curve indicates the Pareto frontier formed by AOrchestra across different model routing choices.